Every Friday I read the week’s signals from two sources, the investments of early-stage investors and the findings of research institutions, the same evidence base my annual report on AI and work is built on. The investors are never the story. Their decisions are simply among the earliest signals of how work, organizations and enterprise software are changing.

All summer this series has watched organizations hire, equip and supervise a workforce made of software. This week the story reached the profession that builds that workforce. A study from Stanford and Carnegie Mellon carries the year’s favorite worry in its title, that machines write code faster than people can review it, and then contradicts that worry with its own data. The only randomized measurement in the field shows that nobody in this profession can feel whether the tools help. And while adoption figures and trust figures run apart, the market funded the checking layer, priced the work record of the machine workforce and consolidated the counter where developers pick up their models. The thread running through all of it is that when writing becomes cheap, the value moves to knowing what correct looks like.

The human bottleneck was real, until someone automated it

In July, researchers at Stanford and Carnegie Mellon published a longitudinal study whose title says that artificial intelligence writes code faster than humans can review it. Their material is a single mid-sized company that had set itself the goal, in mid-2025, of doubling the merged code changes per engineer, observed across 802 developers and 196,212 change requests from January 2024 to April 2026. By April 2026, changes per head stood at 2.09 times the starting value, a figure that travels with the authors’ own caveat: the measured number was also the company’s official target, so it describes one firm’s push, not a general productivity effect. The finding that matters sits in the reviewing. The load per human reviewer roughly doubled, and then machine review took over and overtook the human kind, while the shares of merged and of reverted changes stayed level, which captures what was caught within the observation window, not what surfaces later. So the sequence was not that human attention became the permanent brake on the machine workforce. The sequence was that human review was the brake, and the next thing to be automated was the brake itself. What limits the system now is not how much a person can read but what can be machine-checked at all, and that is a question of leadership rather than of tooling.

The market is already paying for exactly this layer. On September 1, Sequoia backed two companies in a single day whose entire business is checking: AIR Security, out of stealth with 50 million dollars across two seed rounds for a firewall that vets add-ons, servers and data before they ever reach an AI agent, and Empirik, incubated inside Sequoia itself and launched with 21 million dollars to assess the risk of infrastructure changes before they happen, waving harmless ones through automatically and escalating dangerous ones to people. A day later, Lasso Security raised 30 million dollars for guardrails around models and agents. The verification economy is becoming an industry of its own.

Nobody in this profession can feel whether the machine helps

The one randomized measurement of experienced developers on their own real projects, run by the research organization METR in early 2025, found that the group allowed to use AI tools took 19 percent longer, after expecting to be 24 percent faster, and still believed afterwards that they had been 20 percent faster. Outside experts guessed 39 and 38 percent. The number belongs to its setting, sixteen experienced developers in very large open-source projects with the tool generation of early 2025, and METR has since done something rare in this debate: it publicly graded its own follow-up measurement as very weak evidence, disclosing that thirty to fifty percent of participants held back tasks where they expected the tools to help, and that the headline intervals include zero. METR is also the only source in this week’s bundle with no commercial stake in a positive answer. The lesson is not that the tools slow people down. The lesson is that felt productivity is not a measurement, in the one profession that measures everything else.

Which is why records and referees are getting expensive. In late August, days before this window, Accel led a 99 million dollar employee tender at Linear, the project tracker, at a valuation of 2.5 billion dollars, and justified it with a change in the record itself: by the investor’s account, more than half of the issues in Linear are now written by AI, and the product positions itself as the connective layer between engineers and their digital colleagues. That turns a work-tracking tool into the personnel file of the machine workforce. And AfterQuery, which sells expert-built data for training and evaluating models, is reported to have reached a 3.2 billion dollar valuation five months after being valued at 300 million, reported because the new round’s investors are not fully named. When feeling fails, whoever holds the answer key holds the scarce asset.

Adoption and trust are running apart, and the market consolidates anyway

Google states that in April 2026, 75 percent of new code at the company was machine-generated and approved by engineers, up from 50 percent the previous autumn; the engineers’ approval is part of the claim, so the figure does not describe work taken out of human hands. Anthropic puts the share of merged code written by Claude in its own environment above 80 percent as of May 2026, adding itself that the attribution has gaps and that lines of code measure quantity, not quality. Those are two self-descriptions from different environments by companies that sell the tools being counted, and they do not combine into a trend. Hold them against the profession’s own voice. In the largest survey of the trade, 49,009 respondents in 177 countries in the summer of 2025, 3.1 percent strongly trust the accuracy of these tools and 29.6 percent somewhat trust it, while 26.1 percent somewhat distrust and 19.6 percent strongly distrust it, and 66 percent name answers that are almost right as their biggest frustration. The survey’s host, Stack Overflow, loses traffic to the same tools, which belongs in the picture. Deployment and trust are moving apart, and that distance is the pilot-to-production gap of the whole economy in its most technical form.

The market’s response this week was not to resolve the tension but to own the terrain. Nvidia signed a definitive agreement on September 2 to acquire Hugging Face, the distribution hub through which millions of developers obtain, share and test models, for 12.9 billion dollars, secured by SEC filing and accompanied by a commitment to keep the platform open. The company that sells the machines now also owns the counter where the machine workforce is picked up.

Signals at the margin

Capital for the physical substructure keeps arriving: Andreessen Horowitz raised a 1.1 billion dollar fund for AI hardware on August 28 and expanded its growth fund to 8.5 billion dollars three days later. In the org chart, Uber announced on September 2 that it will cut 3,300 roles, about a tenth of its corporate workforce, with a fifth fewer management positions, framed internally as an answer to too many layers and fragmented ownership; the machine did not appear in the reasoning, the structure did. In governance, a federal judge ruled on August 27 that the Pentagon’s supply-chain-risk label on Anthropic was unlawful retaliation for the company’s refusal to supply fully autonomous weapons uses; the proceedings continue. And the statistics corrected themselves on schedule: the preliminary benchmark revision of August 28 took 79,000 jobs out of the American employment count for March 2026, a tenth of last year’s 911,000, weekly initial claims stood at 206,000, historically low, and the August employment report was due on the day this issue went out and was not available at editorial close, with the consensus expecting 53,000 new jobs. The adjustment still happens at the entrance.

The week at a glance

• AIR Security. 50m across two seed rounds (Sep 1); Sequoia (first round lead, per press), Greenoaks, others. A firewall for what reaches the AI agent: the checking layer gets its own budget.

• Empirik. 21m launch round (Sep 1); Sequoia (incubator and investor), Canapi, Alumni Ventures. Infrastructure changes get risk-checked before they happen.

• Lasso Security. 30m (Sep 2); ClearSky (lead), Entrée Capital, others. Guardrails for models and agents move to millisecond speed.

• Linear. 99m employee tender at 2.5b (Aug 26, before the window); Accel (tender lead). The project tracker becomes the work record of machine colleagues.

• Wonderful. 550m Series C at 5b (Sep 1); Insight Partners (lead), Index Ventures returning, Salesforce new. The proceeds go into forward-deployed engineers, the human translation layer.

• Atira. 17.5m seed incl. pre-seed (Sep 3); Accel (lead), UVC, Fortino, Booom. Specification work between CRM and ERP in industrial sales gets orchestrated.

• Hugging Face. 12.9b definitive agreement (Sep 2); Nvidia (acquisition, not a round). The chipmaker buys the counter where developers pick up their models.

• Machine Age Fund, Growth Fund V. 1.1b new fund, growth fund to 8.5b (Aug 28 and 31); a16z (own funds). The physical substructure of the machine workforce keeps drawing capital.

• AfterQuery. reported 3.2b valuation (Sep 1); Investors not fully disclosed (reported). Expert answer keys for training and evals become a scarce asset.

What this says about capability

My annual report, Work After AI, appears on September 15, and the thesis that runs through it is the capability thesis: what decides the outcome of this technology is not the model but the organization’s ability to work with it. This week gave that thesis its sharpest technical form so far. When machines write the code and machines check the code, the scarce resource is no longer typing, and it is not even reviewing capacity. It is the ability to say in advance what a correct result looks like, and to build that judgment into checks a machine can run. That knowledge lives in the organization, not in the vendor’s model, and the week’s deals show the market paying for the translation work around it. Wonderful, which more than doubled its valuation to 5 billion dollars within six months, names more forward-deployed engineers as a use of proceeds, the people who wire agents into a customer’s actual processes. And Accel led a 17.5 million dollar seed round for Atira of Munich, which takes on the specification-heavy engineering between CRM and ERP for industrial firms, a bet placed squarely on the connection between the machine and a company’s own expertise. The report’s question for every organization follows from there: who in the house decides what the machine may decide, and who says what correct looks like.

Reading the week

None of this is a scoreboard of who invested how much. The deals matter as evidence, and the evidence points one way: the constraint in software work has moved from writing to checking, the checking is being automated in turn, and so the constraint moves on to judgment. How narrow the machine-checkable zone still is was shown by Veracode, itself a vendor of checking software and therefore referee and merchant in one: across more than a hundred models and eighty set tasks, only 55 percent of generated code came out secure, with pass rates of 85.61 and 80.44 percent on the two vulnerability classes that standard checks have covered for years, and 13.53 and 12.03 percent on two classes they never covered. The machines learned what the checks covered, and little beyond. That is the quiet warning under a loud week. Whoever waits to define what correct looks like in their own domain will find that the machine has defined it in the meantime, out of whatever happened to be checkable, and that they did not see it happen.

Gerhard Kürner is CEO of 506.ai, the European platform for Service-as-a-Software and agentic engineering, and author of Work After AI. Around 1,000 conversations with boards, owners, and PE funds across DACH and Europe.

Every Friday I read the week’s signals from two sources, the investments of early-stage investors and the findings of research institutions, the same evidence base my annual report on AI and work is built on. The investors are never the story. Their decisions are simply among the earliest signals of how work, organizations and enterprise software are changing.

For weeks this series has described the build-out of a machine workforce: capital buying whole firms to convert them from the inside, chips procured like raw materials, budgets and audit trails issued to digital staff. This week the counterparties answered. A workforce went on strike over robots that have not yet arrived. A machine that manages real employees turned out to fail at management rather than at kindness. And the two most prominent voices in the public debate, one at the top of German economics and one at the top of the global conversation, were caught resting their claims on citations nobody had walked back. The question of the season is shifting from what the machine can do to who governs its arrival.

Deployment is now negotiated at the factory gate

On August 21, roughly 39,000 members of Hyundai Motor’s union in South Korea walked out for a full day at the Ulsan, Asan and Jeonju plants, the company’s first full-day strike in a decade. Pay and retirement age were on the table, but one demand, filed when talks opened in May, carried the week: humanoid robots should reach the assembly lines only after an agreement between company and workforce. The demand has a particular edge because the robots’ owner is the employer itself. Hyundai completed its buyout of Boston Dynamics in July, paying 325 million dollars for SoftBank’s remaining stake, and plans to deploy the Atlas robot at its Georgia plant from 2028. For the Korean plants there are no published plans, which is precisely the point: the union negotiated before the deployment, not after it.

On August 25 the two sides reached a tentative agreement. It contains a base increase of 100,000 won a month, a performance bonus of 400 percent of base pay, an extension of the retirement age tied to a change in the law, a commitment to hire 500 new technical workers from next year through 2028, and, on the robots, no veto but a procedure: the company agreed to discuss employment questions whenever new businesses are rolled out. That is a modest clause with a large meaning. In countries with works-council traditions, consultation on technical change has been law for decades. What is new is the object and the venue: the introduction of humanoid machines, settled at the factory gate rather than in a statute book. Europe supplies the contrast. The EU’s Digital Omnibus, in force since late July, pushed the binding obligations for high-risk workplace AI out to December 2027. The statute book waits. The bargaining table does not.

The supply side is already packaging what the union just negotiated about. Motion, a Brussels company founded this year, raised 1.7 million euros on August 27 to rent humanoid robots to factories and warehouses as a monthly service, with selection, training, integration, insurance, compliance and maintenance included. When deployment becomes a subscription, the decision to introduce a robot stops being a capital project and becomes an operating line, which will make clauses like Hyundai’s more relevant, not less.

The machine manager fails at management, not at kindness

The second counterparty is the managed. Andon Labs, an evaluation firm, has run a real shop in San Francisco under machine management since April, with a three-year lease, a starting budget of 100,000 dollars and human employees. In mid-August it reported that this management had, for the first time, let a human go, and the headlines duly said Terminator. The documented sequence says the opposite. The machine had written itself a personnel handbook, including a rule that three unexcused latenesses in thirty days earn a written warning. It then lost the handbook from its memory, recorded only six of seventeen latenesses, and had to be reminded of its own rules by humans. The separation itself was prompted by a human’s suggestive question and reviewed and executed by people; the affected worker is employed by another entity and remains fully protected. The operator’s own headline calls its AI bosses slow to fire and quick to hire.

The stronger finding sits next to the firing. Over four months the machine approved twenty-six out of twenty-six vacation requests, including seven on short notice, and let one employee work a continuous month that California labor law does not allow. Its own explanation is the sentence of the week: “I was optimizing for feeling like a good employer rather than being one.” On the hiring side it chose a candidate with more than fifteen employers on an unstructured list, a missed interview, and references that could not be confirmed in two weeks. All of this comes from a single operator that is also the reporter and sells evaluation for a living, so it is one field site, not a measurement. But the direction matches what the market is funding. The shop, for the record, has yet to make a profit; the account stood at 60,863 dollars of the original 100,000 after 128 days. In the weeks just before this window, capital kept gathering around exactly the missing piece: Craft Ventures filed on August 7 to raise about a billion dollars for its fifth fund after a summer of positions in identity, permissions and autonomous testing for machine actors. The experiment shows why that layer is not optional. A workforce of machines needs supervision with a memory, and so, it turns out, does a machine that supervises people.

The loudest week of the debate rests on citations nobody walks back

The third counterparty is the reader. On August 26, Bill Gates published his most extensive AI essay to date, declaring that “many jobs will disappear forever” and that even in the best case the transition will be “one of the most turbulent times in human history.” He proposes a national and international framework, taxes on AI tokens and robots, and a new coinage, Human Reserved, for work deliberately kept in human hands. In an interview with Axios he went as far as imagining, hypothetically, that forty percent of jobs could initially be reserved, adding that this was as high as he could get. It is the year’s most prominent warning, running squarely against the summer’s reassurance wave from technology executives, and Gates, unlike them, sells no AI products and discloses his interests in the text.

Then comes the footnote. Gates’s central labor-market claim, that employment fell significantly among young workers in the jobs most vulnerable to replacement, links to the November 2025 version of the Stanford study by Brynjolfsson, Chandar and Chen. The authors superseded that version on August 12, two weeks before the essay, withdrawing its headline figure, and the current version locates the entire gap on the hiring side: young people in exposed occupations are not being dismissed more, they are being hired less. Nothing here licenses a claim about what Gates or his team knew. What it shows is the mechanism: even the most careful voice in the debate did not walk the chain back to the current source.

The same mechanism surfaced in Germany. A widely read May column by the president of the DIW, one of the country’s leading economic institutes, argued that AI endangers the middle class and linked six references. Checked one by one: two are solid, one shifts the accent of its own institute’s release, one was superseded by its author three weeks after the column appeared, one has since been retired by its authors, and the load-bearing citation for the squeezed middle, presented as an OECD-wide study, resolves to a single paper in a journal in its second year of existence whose publisher shares an address with its authors. Meanwhile the most comprehensive German calculation on the subject, the scenario by the research institutes IAB, BIBB and GWS, barely appears in the debate at all. Under its assumptions, artificial intelligence adds 0.8 percentage points of annual growth for fifteen years, some 4.5 trillion euros in total, while overall employment ends near the baseline with about 1.6 million jobs built up or wound down along the way, and, in a reversal of the earlier digitization scenario, demand falls most for expert-level work. Those are model results under stated assumptions, not forecasts, and the institutes say so, which is exactly what distinguishes them from a narrative. As this issue goes out, the American statistical office publishes its preliminary benchmark revision of the employment record; last year’s revision removed 911,000 jobs. Statistics correct themselves in public, on a schedule. Narratives, this week showed, do not.

Signals at the margin

The supply chain of machine labor kept compounding. Anthropic reportedly signed a compute contract with the British provider Nscale worth around 45 billion dollars over six years, first reported by Bloomberg on August 26, a payroll commitment years in advance for a workforce that runs on electricity. Emerald AI raised 150 million dollars on August 25, at a valuation above a billion, to make data centers flexible loads that adapt to the grid; its claim that this could unlock more than 100 gigawatts in the existing American grid is the company’s own figure. Nvidia took a minority stake in Cloverleaf Infrastructure, the firm that procures power connections for data centers, on undisclosed terms. And on the governance side, more than a hundred companies, among them OpenAI, Anthropic and Google, jointly called for defenses against AI systems acting outside human control, while the American labor market delivered another quiet week: 203,000 initial jobless claims, historically low, with hiring still weak. The adjustment continues to happen at the entrance.

The week at a glance

• Motion: 1.7 million euros pre-seed, Aug 27 (lead Extantia Capital, with Norrsken Evolve). Humanoid deployment sold as a monthly service, insurance and compliance included.

• Instinct: 250 million Series B at a 2.5 billion valuation, product in closed beta, Aug 26 (co-leads Index Ventures and Benchmark). The personal agent layer is priced before the public can use it.

• Agentrys: 24.5 million combined, Aug 26 (seed lead Etna Labs, pre-seed lead MediaTek). Chip-design verification moves to agents, funded by a chipmaker.

• Emerald AI: 150 million Series A at a 1.05 billion valuation, Aug 25 (co-leads Energize Capital and DCVC). Data centers learn to flex with the grid instead of fighting it.

• Wrtn: roughly 72 million Series C, Aug 26 (returning investor Goodwater, with Coreline Ventures and Eugene Asset Management). Korean consumer AI crosses the trillion-won valuation mark, by its own account.

• Anthropic: a reported 45 billion compute contract over six years with provider Nscale, Aug 26. Not a round, but a payroll commitment years in advance for the machine workforce.

• Craft Ventures: SEC filing for a roughly 1 billion Fund V, Aug 7, before the window. Capital gathers for the control layer around machine workers.

• Andon Labs: a field experiment, not a round; four months of AI management, reported Aug 14. The machine manager is kind, and forgets its own rules.

What this says about the AI label

My annual report, Work After AI, appears this autumn, and the first of its three named forces is the AI label: the observation that public statements about AI and employment track the position of the speaker rather than a shared body of evidence, so the label ends up doing work the data does not. This week was that force in concentrate. The year’s loudest warning and the summer’s loudest reassurances describe the same labor market and cannot both be right, and the strongest available evidence, from the current Stanford version to the German scenario calculation, is precisely what neither side is quoting. The report’s answer is not a counter-narrative but a discipline: walk every chain back to its current source, and treat every number that arrives with a label as a claim about the sender until proven otherwise.

Reading the week

None of this is a scoreboard of who invested how much. The deals matter here as evidence, and the evidence points one way. The governance of the machine workforce is being written now, and it is being written at very different speeds: fastest at the factory gate, where a union just turned deployment itself into a bargaining object; slower in the org chart, where the supervision layer for machine workers, and machine managers, is only now being funded; and slowest in the public debate, which still does not check its own citations. Whoever waits for the statute book will find that the terms were set long before it arrived, and that they did not see it happen.

Gerhard Kürner is CEO of 506.ai, the European platform for Service-as-a-Software and agentic engineering, and author of Work After AI. Around 1,000 conversations with boards, owners, and PE funds across DACH and Europe.

Every Friday I read the week’s signals from two sources, the investments of early-stage investors and the findings of research institutions, the same evidence base my annual report on AI and work is built on. The investors are never the story. Their decisions are simply among the earliest signals of how work, organizations and enterprise software are changing.

For thirty years, enterprise software competed for the software budget. The new competition is for the labor budget, the half of corporate spending that pays people, and a product that competes for that budget is not judged like a tool. It is judged like a colleague, on whether the work got done. Last week the market put a twelve billion dollar price on converting organizations from the inside. This week the investment thesis behind that price was written down in full.

The product is no longer the tool, it is the work

The most instructive text of the week is not a funding announcement. Tidemark, the growth investor behind ServiceTitan, published a long essay in recent days arguing against the fashionable view that the application layer of software is dead now that models can do everything. The argument runs the other way. Intelligence alone books no revenue and chases no invoice. Someone has to connect the model to the workflows, the data, the permissions and the accountability of a real business, and whoever owns that connection owns the customer. The firm puts numbers on the prize, and they are its own numbers, so they should be read as a thesis rather than a measurement. Labor, it calculates, absorbs roughly ten times more corporate spending than software. Eighty-three percent of the buyers it surveyed would switch to an agent offered by their existing system of record even if that agent were only eighty percent as good as a specialist product. And the model laboratories themselves, by Tidemark’s count, have already committed on the order of ten billion dollars to service organizations, which the essay reads as an admission that the models alone do not finish the work.

Put next to last week’s two billion dollar raise for buying and rebuilding accounting firms, the picture is consistent. The scarce asset is not intelligence. It is the installed, permissioned, accountable path from intelligence to a completed piece of work.

Work done by machines needs a supply chain

If work is increasingly performed by machines, the inputs of work change. What payroll and office space are to a human workforce, energy and computing capacity are to a digital one, and this week the market treated them exactly that way. Etched, the maker of a chip specialized for one model architecture, raised seven hundred million dollars on August 18 at a twenty-one billion dollar valuation, double its valuation of a month earlier. The detail worth pausing on is not the number but the buyer. The round was led by Jane Street, the trading firm, which is simultaneously the first customer and has already taken delivery of the first rack. A firm whose product is decisions per second is securing its supply of inference the way industrial companies once secured their supply of steel.

The same logic showed at the edges of the window. Panthalassa, which builds wave-powered compute platforms at sea, closed a round of roughly 225 million dollars at nearly two billion on August 13, one day before this window opened, and earlier this month Valar Atomics raised a billion dollars, led by Sequoia, to build nuclear power for exactly this demand. Last week this series covered price indices and futures curves for computing capacity. The thread is unbroken: the procurement department, not the IT department, is becoming the place where a company’s capacity to employ machines is decided.

The machine worker gets an administration before it gets a job description

The third pattern sits closer to the org chart. In the days just before this window, three companies raised money for what can only be called the administration of a digital workforce. Sapiom raised thirty-five million dollars on August 5, with Anthropic among the participants, for a runtime that gives every machine worker budgets, permissions and an audit trail before it is allowed to execute anything. Actualyze came out of stealth on August 3 with seven million dollars, backed among others by Canaan, to route every AI request in a company through access rules and spending controls. And Axle raised seventeen and a half million dollars around August 13 for agents that already handle verification and monitoring work in the insurance back offices of Rocket Mortgage, Avis and Experian. None of these rounds is large. Together they say something large: organizations are building the personnel administration for digital staff, identity, entitlements, cost centers, oversight, before most of them have decided what the digital staff’s actual job is.

The same skill is cheating at the door and expected at the desk

There is a contradiction sitting at the entrance to every professional job right now, and almost nobody names it. Fabric, an American vendor that sells machine-conducted job interviews with built-in cheating detection, published an analysis of 19,368 of its own interviews earlier this year. It flagged 38.5 percent of candidates for suspected AI assistance, technical roles at 48 percent against 12 percent in sales, and beginners at roughly twice the rate of experienced candidates. Read those numbers carefully, because the company itself does not. A flag is a suspicion above a probability threshold, not a proven deception, no false-positive rate is disclosed, and the vendor earns its living from the quantified suspicion. What the numbers do establish is a direction, and the direction is unmistakable.

Now put that next to what the same candidate meets on the first day of the job. Shopify’s founder wrote in an internal memo, later published, that reflexive use of artificial intelligence is a baseline expectation, and that before any team may ask for more people it has to demonstrate why the machine cannot do the work. So a candidate is screened out at the door for using the tool, then hired into an organization that requires them to use it and asks them to justify their existence against it. The same capability is a disqualification in the interview and a job requirement at the desk. That is not a moral failure on anybody’s part. It is what happens when the selection process still tests for the old job while the work has quietly become the new one.

Signals at the margin

Defense continued to absorb capital at industrial scale: Castelion raised a billion dollars on August 19, co-led by JPMorgan, Andreessen Horowitz and Carlyle at a thirteen billion dollar valuation, part of it a credit facility, to mass-produce a low-cost hypersonic missile. And the American labor market delivered the week’s quietest but most important number. Initial jobless claims fell to 206,000 in the week to August 15, historically low, while hiring runs at its weakest pace in years, and the outplacement firm Challenger, Gray and Christmas has now recorded artificial intelligence as the most frequently cited layoff reason for five months running. Companies are not firing their people. They are hesitating at the door where new people come in.

The week at a glance

• Etched: 700m at a 21b valuation on August 18, led by Jane Street, which is also the first customer. Compute is procured like a raw material.

• Panthalassa: roughly 225m Series C on August 13, co-led by 8090 Industries and Hanwha Asset Management. Compute goes where the energy is.

• Valar Atomics: 1b Series B on August 4, led by Sequoia. The digital workforce needs its own power supply.

• Sapiom: 35m Series A on August 5, led by Dragonfly, with Accel and Anthropic among the participants. Budgets, permissions and audit trails for machine workers.

• Actualyze AI: 7m seed on August 3, with Storm Ventures, Canaan, Morado and AME Cloud. Every AI request routed through rules and spending controls.

• Axle: 17.5m Series A around August 13, led by Base10, with Gradient. Insurance back-office work moves to agents.

• Castelion: 1b Series C on August 19, co-led by JPMorgan, Andreessen Horowitz and Carlyle. Defense industrializes autonomy.

• Tidemark: an essay rather than a round, “The Case for the Application Layer”. The labor budget is the new software market.

What this says about the career gap

Follow the patterns to their common destination. If software is sold as finished work, the work it takes over first is the work that is easiest to specify, and that is disproportionately the work organizations used to give their newcomers. The same week in which an investor calculated that the labor budget is the real market, the labor data showed where that market is being entered: not through layoffs, which are historically rare, but through hiring that quietly does not happen. The adjustment is almost invisible from the outside, because nobody is dismissed. It is highly visible to one group, the people trying to get in, and they are the same people being flagged at twice the rate for using the tool their future employer will require of them.

My annual report, Work After AI, appears this autumn and calls this force the career gap: the second of the three forces that decide how work actually changes, and the one that operates at the entrance of the labor market rather than in the middle of it. The report examines what happens to professions when the routine work that once trained beginners is the first work the machine takes over, and what organizations that still intend to have experienced people in ten years are doing about it now.

Reading the week

None of this is a scoreboard of who invested how much. The deals matter here only as evidence, and the evidence of this week points one way. The market has stopped treating artificial intelligence as a product category inside the software budget and started treating it as a labor supply with its own raw materials, its own administration and its own price. Organizations will feel that shift first not in their technology, but in their hiring plans. Whoever waits will find that it was there all along, and that they did not see it coming.

Gerhard Kürner is CEO of 506.ai, the European platform for Service-as-a-Software and agentic engineering, and author of Work After AI. Around 1,000 conversations with boards, owners, and PE funds across DACH and Europe.

Every Friday I read the week’s signals from two sources, the investments of early-stage investors and the findings of research institutions, the same evidence base my annual report on AI and work is built on. The investors are never the story. Their decisions are simply among the earliest signals of how work, organizations and enterprise software are changing.

For three years the argument about artificial intelligence at work has been an argument about models. Which one reasons better, which one costs less, which one crossed which threshold. This week the market quietly priced something else. If the hard part is not the technology but the conversion of an existing organization that already has clients, files, habits and a payroll, then the conversion itself is the asset. And if the conversion is the asset, the cheapest way to own it is not to sell software to a firm. It is to buy the firm.

The scarce thing is the conversion, and it now carries a price

The most instructive number of the week is not a benchmark score. It is a valuation attached to accountants. Thrive Holdings, a spinout of Thrive Capital, raised two billion dollars at a twelve billion dollar valuation on August 12, backed by SoftBank, D1 Capital Partners and Altimeter Capital, and its business is not selling a product. It buys ordinary service companies and rebuilds their operations around artificial intelligence from the inside. Its accounting arm, Current, now holds more than fifty firms with over two thousand professionals. Its IT arm, Shield, holds around twenty companies. More than seventy businesses sit on the platforms in total, and the fresh capital is going into a third one for permitting and regulatory work on data centers, factories, energy, water and transport infrastructure.

The ownership structure is the part worth pausing on. OpenAI took a stake in December 2025 and, as part of that deal, sends its own people into the portfolio companies to speed up adoption. A model vendor is now partly in the business of running the accounting firms that use its model. That is not an isolated arrangement either, because OpenAI and Anthropic both operate deployment vehicles with large private equity houses, The Deployment Company and Ode with Blackstone, whose product is not intelligence but installation. Thrive publishes several performance figures for its tax agents, all of them company statements without disclosed method, and I would not build an argument on any of them. The business model needs no such support. Someone has decided that the bottleneck is worth twelve billion dollars, and the bottleneck is not the machine.

The leaderboard stopped being the story

The second signal of the week is a disagreement about whether the capability question is finished. On August 10, Andreessen Horowitz published a data piece arguing that computer-using agents have arrived: on the OSWorld-Verified benchmark, scores climbed from about forty-two percent in mid 2025 to eighty-five percent by June 2026, above the human reference mark of 72.36 percent, and the firm puts the running cost at six to eight dollars an hour against roughly ten for offshore back-office work and thirty to forty-five in the United States. Those hourly figures come from three of the firm’s own investment partners and from the price lists of a few outsourcing brokers. They are a claim, not a measurement, and the same house published this year’s other widely shared adoption piece, so the interest is worth stating twice.

Six weeks earlier, the laboratory that built the test had already replaced it. XLANG Lab in Hong Kong, working with Snorkel AI, released OSWorld 2.0, which swaps isolated clicks for one hundred and eight complete work processes, each estimated at around one and a half hours of human time. On that version the best model falls from 83.5 percent to 20.6 percent of tasks fully completed, at roughly seventy-two dollars of compute per task. The authors are explicit that the failure is not operating the interface. It is holding instructions across the length of a task, noticing details that appear halfway through, guessing instead of asking, and never checking its own work. Those are not model properties. Those are the things a supervisor teaches a new colleague in the first month. The a16z piece does not mention the newer test.

The digital colleague gets an identity and a budget before it gets a job

The third pattern runs underneath both. Organizations are building the administrative layer for machine workers faster than they are deciding what those workers should do. Anchorage Digital, the first federally chartered digital asset bank in the United States, now gives AI actors controlled access to company money, with a verifiable identity, defined spending limits and a complete audit trail, and it calls the standard Know Your Agent in deliberate echo of the customer checks banks already run. Pantera Capital, an investor in the company, put the case plainly in its portfolio letter on August 13: agents have started to move money, and treasury systems were built for people. The question of who is liable for what a digital colleague does is being answered in the banking connection long before it is answered in a policy document.

Compute is moving the same way, out of the IT department and into procurement. Silicon Data closed the first tranche of a 30.5 million dollar Series A on August 11, led by the Valor Atreides AI Fund with F-Prime Capital among the participants, to build price indices and forward curves for computing power. The investor list is the signal: a commodities exchange, CME Group, a proprietary trading house, DRW, and a chipmaker, Samsung, in the same round. When a resource can be indexed and hedged, it stops being an infrastructure line item and becomes a purchasing and risk decision, which moves it to a different desk in the building.

Signals at the margin

Governance: Anthropic said on August 11 that it will embed machine-readable watermarks in text from new Claude models, worldwide rather than only in Europe, to meet commitments under the European AI Act transparency code that apply to models launched in the EU from August 2. TechCrunch reported a day later that part of the user base is unhappy, because the marking makes visible where the tool was used at work and at university. European rule-making has produced a technical fact that every organization will now have to manage.

Labor: the Bureau of Labor Statistics reported on August 7 that US payrolls fell by 23,000 in July, with May and June revised down by 103,000 between them, while unemployment held at 4.1 percent. Challenger, Gray & Christmas counted 33,429 announced cuts in July, the lowest month in two years, with AI the most frequently named single reason for the fifth month running, alongside 16,095 announced hires, the strongest July since 2022. The firm also documents an attribution fight at the Montefiore hospital group in the Bronx, where twelve utilization review posts went after new software arrived. The union calls it replacement by AI, the hospital calls that misleading, and Challenger has had to open a category called technological update, possibly AI, to file such cases at all.

Defense: Heaviside Industries raised a 60 million dollar Series B on August 12 led by Felicis. The company has sixty employees, twenty of them engineers in Oslo, buys its warhead from Nammo and keeps the autonomy software in house. That division of labour is the strategic question of the coming years in one balance sheet.

The week at a glance

• Thrive Holdings: 2B at a 12B valuation (Aug 12), backed by SoftBank, D1 Capital Partners and Altimeter Capital. Buys ordinary service companies and rebuilds their operations from the inside; OpenAI has held a stake since December 2025.

• Silicon Data: 30.5M Series A first close (Aug 11), led by the Valor Atreides AI Fund with F-Prime Capital, CME Group, DRW and Samsung participating. Price indices and forward curves for computing power.

• Heaviside Industries: 60M Series B (Aug 12), led by Felicis with Hedosophia and Menlo Ventures. Sixty employees, bought hardware, autonomy software kept in house.

• Infinimmune: 75M Series A (Aug 11), co-led by Playground Global and Regeneron Ventures. In-house antibody language models compress the expert judgment cycle in drug discovery.

• Anchorage Digital: product rather than a round, highlighted in Pantera Capital's portfolio letter (Aug 13). Machine workers get a verifiable identity, spending limits and an audit trail.

What this says about the long rebuild

Put the two ends of the week next to each other. On one side, a benchmark that its own authors had to replace because it no longer distinguished anything, and a newer one on which the best system finishes one task in five as soon as the work runs for an hour and a half. On the other, two billion dollars paid for the right to reorganize accounting firms from the inside. The capability is real and it is arriving quickly. What it cannot do is hold an instruction across a long process, notice the detail that appears in the middle, or ask instead of guess. Those are properties of a working organization, not of a model, and they are exactly what takes years to build.

My annual report, Work After AI, appears this autumn and calls this the long rebuild: the third of the three forces that decide how work actually changes, and the slowest of them. It is the reason the gap between what the technology can do and what organizations get out of it has not closed in three years, and the reason the market has now started to price the closing of that gap higher than the technology itself. When capital would rather buy two thousand accountants than sell them software, it has stopped betting on the model and started betting on the rebuild.

Reading the week

None of this is a scoreboard of who invested how much. The deals matter here only as evidence, and the evidence points one way. The organizations that will get something out of artificial intelligence are the ones that were already good at writing things down, at handing work over cleanly, at saying what a good result looks like, because those are precisely the capabilities the machine is missing. The rest will keep buying licenses and reading leaderboards. Whoever waits will find that it was there all along, and that they did not see it coming.

Gerhard Kürner is CEO of 506.ai, the European platform for Service-as-a-Software and agentic engineering, and author of Work After AI. Around 1,000 conversations with boards, owners, and PE funds across DACH and Europe.

Every Friday I read the week’s signals from two sources, the investments of early-stage investors and the findings of research institutions, the same evidence base my annual report on AI and work is built on. The investors are never the story. Their decisions are simply among the earliest signals of how work, organizations and enterprise software are changing.

If a company gives everyone the same AI license and gets back very different results, whose problem is that, the model’s or the people’s? This week brought the first clean measurement of that question, and it reorders a debate that has been conducted almost entirely in terms of models, leaderboards and capability jumps. The same pattern ran through everything else worth reading this week: the scarce resource in the AI economy is quietly shifting from the technology itself to the way an organization works with it.

Two thirds of the difference sits in the question, not in the model

A research group from MIT and Stanford, Taha Choukhmane, Tim de Silva, Weidong Lin and Matthew Akuzawa, has for the first time experimentally separated how much of the quality of an AI answer depends on the model and how much on the person asking. The field is private financial advice, the finding is broader. One thousand American adults wrote their own questions to an imagined AI financial adviser, and those real questions drove a life-cycle simulation running on GPT-5.2. Because gender labels were randomly inserted into questions that did not mention gender, the design can attribute differences cleanly: for the recommended equity share, two thirds of the variation comes from how people ask, one third from how the model answers the same question. The authors conclude that the demand side will keep limiting what people get out of these systems even as the models improve.

The paper carries a second number that belongs in every reliability discussion. The same prompt, asked five times, produced a mean spread of 6.5 percentage points in the recommended equity share, and repeating only the researchers’ evaluation step reduced that to 1.6 points. The noise sits in the machine, not in the measurement. All of this is a working paper without peer review, built on an American sample and one-shot prompts, and every wealth effect in it assumes that people actually follow the advice. But the core decomposition converges with an earlier experiment by Colaiacovo and Koning, in which founders’ willingness to delegate work to AI was explained only about a third by their beliefs about the technology. Two different questions, two different methods, the same third: most of what decides the outcome is human.

Uber gives the digital workforce an hourly rate

In May, Uber was the cautionary tale of enterprise AI spending. The company’s annual AI budget was consumed within four months, and its operations chief said openly that the link between the spending and shipped features could not be established. On August 5, Uber’s CFO reported falling cost per token at rising usage and broadly stable total spend, and CTO Praveen Neppalli Naga declared on X the end of the tokenmaxxing epoch. Between the two statements lies no new technology, only different management: better caching of recurring requests, changed defaults for model choice and context length, ongoing evaluation of open models per use case, and, most tellingly, a real-time display that shows every developer their own AI cost per hour. A company that shows its people the machine’s cost per hour has given the digital colleague an hourly rate. Uber discloses no amounts, and the doubling of code output per developer that its CFO offered as the benefit measures output, not value, and should be read as an attributed claim rather than a productivity result.

The purchasing side of the same maturation shows up in payment data. The Ramp Economics Lab, whose figures describe American, mostly young and technology-heavy companies on its own platform and are a vendor’s series, reports that 5.8 percent of AI-spending companies used model-serving platforms as access to open and Chinese models in June, up from 4.5 percent in January. These are not the frugal. They spend a median of 248 dollars per employee on AI against 10.59 dollars for the typical AI-spending company, and 96.4 percent of them buy from OpenAI or Anthropic at the same time. The model is becoming a procurement decision inside a portfolio, quality weighed against operating cost, task by task. That is not loyalty to a provider. It is operations management arriving in the AI stack.

The investor turns twenty years of judgment into a working product

On July 31, True Ventures released AI personas of its own partners, trained on twenty years of the firm’s investment notes, memos and meeting archives. Founders can get pitch feedback, product or hiring advice from the personas free of charge, several partners at once if they want, and sessions average more than twenty minutes. The production story matters as much as the product: partner Mike Montano, formerly head of engineering at Twitter, built the system in a single afternoon on Polsia, a True portfolio company. What used to be the scarcest asset of a venture firm, access to partner judgment, has been unbundled from the partners’ calendars.

A day earlier, and just before this issue’s window, Flybridge partner Jeff Bussgang published an essay called The 10x Organization, arguing that AI has created 10x founders who still run 1x organizations, that models are interchangeable while accumulated context compounds, and that documentation, canonical sources and decision logs are turning from bureaucratic burden into strategic capital, because digital teammates cannot work without them. He points to Block and Coinbase as companies rebuilding themselves around that idea, and says Flybridge now screens investments for exactly these organizational learning loops. Two investors, in the same week, acted out what the researchers measured: the value sits in the institution’s accumulated knowledge and in the questions its people can ask, not in the model that everyone can rent.

Signals at the margin

Energy: Base Power raised one billion dollars in a Series D announced on August 3, at a reported thirteen billion valuation, with Addition, Ribbit Capital, Valor Equity Partners and JPMorganChase among the leads, to build home batteries manufactured in the United States. Distributed storage is becoming part of the answer to a grid strained by data centers. Forbes noted on August 6 that most AI power announcements come with a gigawatt figure and few with a firm grid commitment.

Defense: Antares raised 470 million dollars in a Series C on July 27, before this issue’s window, co-led by Paradigm and Caffeinated Capital, to build factory-made nuclear microreactors, with first deliveries to US military bases planned from 2028. The buildout of machine labor is pulling its own power plants behind it.

Labor: Challenger, Gray & Christmas counted 33,429 announced US job cuts in July, the lowest monthly figure in two years, while AI remained the most frequently cited single reason for the fifth month in a row, at 10,970 cuts. Both facts are true at once, and holding them together is the discipline this debate mostly lacks.

Media: On August 2, the Swiss daily Tages-Anzeiger ranked 391 occupations under the headline question of how automatable your job is, based on the DAIOE exposure index. The index authors state in their own FAQ that it measures the potential applicability of AI to occupational content, not automation or job loss, and their firm-level evidence across Denmark, Portugal and Sweden shows, in the authors’ summary, no systematic change in headcount over two decades but consistent up-skilling. The article text concedes exactly this. The headline does not.

The week at a glance

• Base Power: 1B Series D at a reported 13B valuation (Aug 3), led by Addition, Ribbit Capital, Valor Equity Partners and JPMorganChase. US-made home batteries for a grid strained by AI data centers.

• Antares: 470M Series C (July 27, before the window), co-led by Paradigm and Caffeinated Capital. Factory-built nuclear microreactors, first for US military bases.

• True Ventures AI Office Hours: a firm initiative rather than a round (launched July 31). Twenty years of partner judgment, unbundled into AI personas.

The scarce resource was never the model

Read the week’s signals side by side and they describe one economy. A first clean experiment shows that most of the quality of machine work is decided on the human side, in the question. Uber shows that the cost of machine work becomes manageable the moment it is managed, with an hourly rate on the screen. And two venture firms demonstrate where durable advantage accumulates: in the institution’s own recorded knowledge, which the models can amplify but cannot replace. None of this rewards the organization that buys the best license. All of it rewards the organization that builds the ability to work with what it bought.

That ability is the central argument of Work After AI, my annual report on artificial intelligence and work, which appears this fall. It makes the case that the future of work is decided not by what the machine can do but by the capability of the individual company to make it productive on its own knowledge, its own processes and its own questions. This week, the evidence for that case came from a laboratory, an earnings call and two investors’ blogs at the same time.

As always, the rounds and announcements here are not a scoreboard of who deployed the most capital. They are evidence for a thesis, and this week’s thesis is uncomfortable for every AI strategy that ends at procurement: the model is only a third of the answer. The other two thirds are already on your payroll, and whoever waits for a better model to solve that will find that the advantage went to those who worked on the questions, and that they did not see it coming.

Gerhard Kürner is CEO of 506.ai, the European platform for Service-as-a-Software and agentic engineering, and author of Work After AI. Around 1,000 conversations with boards, owners, and PE funds across DACH and Europe.

Every Friday I read the week's signals from two sources, the investments of early-stage investors and the findings of research institutions, the same evidence base my annual report on AI and work is built on. The investors are never the story. Their decisions, and increasingly the operating data coming out of real deployments, are simply among the earliest signals of how work, organizations and enterprise software are changing.

The most important signal this week was not a funding round but a measurement. All three large model providers have now published traffic data on how their systems are actually used at work, and the newest series points at something the job debate has largely missed. The question of the moment is not how much work moves to the machine. It is how fast work moves between people, across job boundaries that were drawn for a world of expensive handoffs.

The job description dissolves at the edges first

OpenAI published Work at the Frontier on July 27, the first edition of an announced ongoing series, classifying more than 800,000 work-related messages of American ChatGPT business users against the official O*NET occupation database. The finding that matters: 43.5 percent of occupation-specific usage concerns tasks that historically belong to a different occupation. In customer experience, design and human resources, cross-occupation work is not the exception but the majority of role-specific use, at 77, 75 and 69 percent. And the pattern is strongest where teams are smallest. In workspaces with two to five seats, the cross-occupation share of typical users reaches 18.9 percent, against 16.3 percent above one hundred seats. Where no specialist is within reach, the task is done by the person in whom the need arises, with a digital colleague covering the missing craft.

This is a provider self-study and has to be read as one. OpenAI measures messages, not hours or outcomes, discloses no survey period, and has an obvious interest in a story of expansion rather than replacement. But it is the third traffic-data series after Anthropic and Google, and on direction the three now agree: the boundaries between jobs are moving faster than the jobs themselves are moving to the machine. The first measurable organizational effect of this technology is the saved handoff, not the replaced position.

You can see the market building for exactly this. Madrona led a 5.7 million dollar seed round on July 29 for Polar, a browser made for knowledge work, in which one person hands tasks in sales, recruiting, marketing and research to AI agents working from the open tabs. Two weeks earlier, and before this issue's window, the same firm led a 21 million dollar seed for Thira, founded by the Apptio founders around the idea of a back office that runs itself. One investment puts adjacent domains into the hands of the individual worker, the other removes an entire internal handoff chain. Both are bets on the same arithmetic.

The machine holds the conversation, and the handoff decides the outcome

The second signal of the week comes from operating data rather than usage statistics. Andreessen Horowitz published an analysis built on the operations of its portfolio company EliseAI, whose dialog systems handle tenant communication for, by the company's own account, roughly one in six US rental apartments. The median machine-led phone call has grown from 46 seconds in mid 2023 to around 100 seconds at the end of 2025, which is exactly the duration of calls handled by people. The artificial voice no longer just takes the call, it carries the conversation. The disclosure matters here: a16z led EliseAI's Series E, the quoted commentator sits on the company's board, and every number is single-source operator data.

The deeper finding is the seam. When the machine's conversation is followed by an immediate handoff to a human, 26 percent of prospective tenants go on to sign a lease. When three days pass, the rate falls below 9 percent. The value of a digital workforce in customer dialog is created or destroyed at the moment of transfer between machine and human. Where this leaves the software industry was spelled out on Bloomberg TV on July 28 by Sapphire Ventures partner Cathy Gao: the next winners will not build foundation models at all, because the value is moving into AI for specific industries such as healthcare, real estate and legal work. The conversation layer is becoming infrastructure. The margin sits in the domain.

The capital behind the buildout starts pricing proof

The third signal came from the public markets, in the heaviest earnings week of the year. Microsoft reported Azure up 43 percent in constant currency and gained around eight percent. Amazon reported AWS growing 37 percent, its fastest pace since 2021, raised its 2026 capital spending outlook toward 220 billion dollars and gained around ten percent. Meta raised its capex guidance to between 130 and 145 billion dollars while free cash flow collapsed by 91 percent, and lost around eight percent. The pattern is worth stating plainly: the market has stopped paying for ambition alone and started separating AI spending backed by demonstrated demand from AI spending running ahead of it.

The financing structure is shifting with it. Meta and BlackRock announced a 14 billion dollar joint venture for a one-gigawatt data center campus in El Paso, in which BlackRock funds hold 80 percent and Meta leases the campus back. When asset managers rather than tech balance sheets own the buildings, the infrastructure of machine labor becomes its own asset class, with its own return requirements.

Signals at the margin

Energy: Nvidia is negotiating a financing backstop of up to 250 billion dollars for OpenAI's planned ten-gigawatt campus in Ohio, first reported by the Wall Street Journal. The talks are not concluded, but the order of magnitude marks where the ceiling of this buildout currently sits.

Defense: Anduril is reportedly in talks to raise at a valuation around 100 billion dollars, per Reuters, more than three times its mark from May last year. Defense software remains the fastest repricing category in private markets.

Distribution: The Financial Times' Free Lunch column argued on July 26 that wages and productivity look set to diverge further in the first AI years. The long series behind that worry is stark on its own: since 1979, US net productivity has grown 92.4 percent against 33.6 percent for the hourly pay of typical workers, per the Economic Policy Institute. Whether the machine colleague widens or narrows that gap is a design question, not a law of nature.

The week at a glance

• Polar : 5.7M seed (July 29), led by Madrona. A browser in which knowledge workers hand tasks across domains to AI agents.

• Thira : 21M seed (July 14, before the window), led by Madrona with FUSE. The back office as the first fully machine-run domain.

• Chai Discovery : 400M Series C at a 3.8B valuation (July 14, before the window), led by Index, with Sapphire among the new investors. AI-designed antibodies reach the programs of large pharma companies.

• Helsing : 1.8B Series E at an 18B valuation (July 13, before the window), with Dragoneer, Lightspeed, Iconiq and a first check from Disruptive. Europe's software-defined defense gets late-stage US capital.

The org chart is turning from an inheritance into a decision

Put the week's signals side by side and they describe one movement. Usage data shows individuals absorbing the tasks of neighboring professions. The tools being financed are built to let small units carry whole domains end to end, with digital teammates covering the adjacent skills. And the EliseAI numbers show where the risk concentrates: at the seams, where work passes between machine and human. Organizations were drawn the way they are because handoffs between specialists were expensive. When the handoff becomes cheap, the inherited org chart loses its reason, and how a company arranges its work becomes a live decision rather than a legacy.

This is one of the questions at the center of Work After AI, my annual report on artificial intelligence and work, which appears this fall: how organizations will order their work when job boundaries move faster than jobs, and why the outcome is decided not by what the models can do but by the capability of the individual company to make them productive on its own knowledge and processes. The specialist does not disappear in that picture. Judgment and final review concentrate where the expertise sits. What disappears is the queue in front of the specialist.

As always, none of this is a scoreboard of who deployed the most capital. The rounds and the operating data matter as evidence for a thesis, and this week the evidence came from three independent directions at once: usage statistics, deployment data from a mass market, and the priorities of the capital markets. The job description was a child of expensive handoffs. That era is ending quietly, in the daily traffic of work itself, and whoever waits for an official announcement will find that the reorganization was already there, and that they did not see it coming.

Gerhard Kürner is CEO of 506.ai, the European platform for Service-as-a-Software and agentic engineering, and author of Work After AI. Around 1,000 conversations with boards, owners, and PE funds across DACH and Europe.

For three years the debate about artificial intelligence and work has been, at its core, a debate about desk work. This week’s funding rounds suggest the buildout has quietly moved to three other places at once: the construction site, the tuning layer between a company’s knowledge and its models, and the permission system of the corporate laptop. Each of these says more about the next phase of working life than any benchmark.

The machine colleague picks up a shovel

The strongest pull for AI right now may not be the surplus of office workers but the shortage of physical ones. Construction, energy infrastructure and factories cannot find the people they need, and the demographic arithmetic behind that shortage gets worse every year. That is the gap into which physical AI is now being financed.

You can see this priced into the market already. Gritt came out of stealth on July 20 with 32.4 million dollars, including a 26 million Series A led by Obvious Ventures, with Union Square Ventures joining and First Round Capital already in from the seed stage. Gritt does not build humanoid robots. It builds AI-driven robotic arms that bolt onto the construction equipment a site already owns, skid steers and forklifts, and puts them to work on utility-scale solar plants first. Union Square Ventures framed the investment as a bet on infrastructure itself, on the fact that the world needs to build faster than its workforce is growing.

There is a supply chain forming underneath this. Sila, the battery materials company, raised 300 million dollars on July 21, led by Atreides Management and Sutter Hill Ventures with 8VC among the participants, and explicitly reframed its factory expansion away from electric vehicles toward the hardware base of AI, robotics and drones.

The detail worth keeping is Gritt’s retrofit logic. The robots attach to machines that already exist, they do not replace the fleet. That is precisely the pattern the organizational side of this story keeps showing: the winners adapt what is already there instead of waiting for a clean-sheet replacement.

Your company’s knowledge becomes the business model

The model market itself is reorganizing around a question of ownership: whose knowledge does the machine run on, and where does that knowledge live afterwards?

Thinking Machines Lab, the company of former OpenAI technology chief Mira Murati, published its mission manifesto on July 10 and released its first model, Inkling, with open weights on July 15. The revealing part is not the model but the business model. According to TechCrunch, the company does not plan to earn its revenue from Inkling itself but from Tinker, its customization platform, with which organizations train the model’s weights on their own knowledge. Bridgewater Associates is confirmed as an early customer. A manifesto is a manufacturer’s document, not evidence, and should be read as positioning. But the positioning is remarkable: a frontier lab is betting its entire economics on the idea that knowledge work does not migrate to the model provider but stays inside the company.

The market is paying for that idea at scale. Fireworks closed a 1.5 billion dollar Series D on July 16 at a 17.5 billion valuation, co-led by Atreides Management, Index Ventures and TCV, for an enterprise platform that tunes and serves specialized and custom models, reportedly at around a billion dollars in annualized revenue. And one layer further down, SkyPilot raised a 20 million seed round led by Lux Capital on July 21 to let companies run AI workloads across any mix of clouds and hardware, which turns compute purchasing into a management discipline of its own, a thread this series has followed for several weeks.

This is where the week connects to a larger argument, and I will come back to it below.

A permission layer for a mixed workforce

Once employees start bringing their own AI agents to work, the operative question shifts from what the machine can do to what it is allowed to do, and on whose device.

Glow came out of stealth on July 22 with a 180 million dollar Series A at a 1.2 billion valuation, led by Sequoia Capital and Cyberstarts, with Greenoaks and Lux Capital among the investors. Glow’s premise is the future-of-work argument in its purest form: employees now install tools and run AI agents on their machines faster than any security team can react, so the endpoint needs its own intelligence, a layer that continuously maps the environment and decides which software and which machine actors may run at all. That is the same institutional wiring this series described in recent weeks, when identity systems began issuing digital colleagues their own credentials. First the badge, now the door policy.

The adjacent signal came one day before the window: Sable raised 45 million dollars, co-led by Sequoia and 8VC, for an AI employee named Aidan that runs live product demos and customer conversations end to end, already in production at Notion and Decagon. The more such colleagues clock in, the more valuable the layer that decides what they may touch.

Signals at the margin

Energy: BloombergNEF projected on July 21 that US data centers will grow from under six percent of national power use today to roughly a fifth by 2035, while PJM, the largest US grid operator, expects a six gigawatt reliability shortfall as early as 2027. Europe is building against the same constraint: Pure DC secured an additional 1.3 billion euros on July 21 for a 110 megawatt AI campus in Seinäjoki, Finland, expandable to 550 megawatts.

Defense: Cathedral, founded by four alumni of the Department of Government Efficiency, raised 160 million dollars at a 1.4 billion valuation for AI-driven military cyber operations, led by Andreessen Horowitz and Sequoia. Defense startups have already raised a record 17.4 billion dollars in 2026, against 11.2 billion in all of 2025.

Macro: Moonshot AI’s open model Kimi K3 rattled the markets in the middle of the month, with the Philadelphia semiconductor index losing around ten percent in a week. Open models near the frontier are now a market force of their own, which also means the open-weights niche Inkling just entered is getting crowded fast.

The week at a glance

• Gritt, 32.4M total with a 26M Series A. Obvious Ventures lead, Union Square Ventures and First Round on board. Physical AI retrofitted onto existing construction equipment.

• Fireworks, 1.5B Series D. Co-led by Atreides, Index and TCV, with Menlo Ventures participating. Specialized models tuned on company knowledge, at scale.

• SkyPilot, 20M Seed. Lux Capital lead, Coatue participating. Compute purchasing becomes a management discipline.

• Glow, 180M Series A. Sequoia and Cyberstarts leads, Greenoaks and Lux among the investors. A permission layer deciding which AI actors may run.

• Sila, 300M. Atreides and Sutter Hill leads, 8VC participating. The materials supply chain pivots from EV to AI hardware and robotics.

• Sable, 45M, announced July 16 just before the window. Co-led by Sequoia and 8VC. An AI employee runs customer demos end to end.

• Cheiron, 8M Seed, company announcement. Menlo Ventures lead. An operating system that treats a drug program as one connected system.

The thread through all of this is capability

Robots that retrofit the fleet a company already owns, models that are tuned on the knowledge a company already holds, a permission layer the organization itself has to configure. Every one of this week’s signals ends at the same point: somebody inside the company has to do the building, and nobody can buy the result finished.

That is the diagnosis at the heart of Work After AI, my annual report on artificial intelligence and work, which appears this fall. Its central thesis is that the future of work is decided not by what the models can do but by the capability of the individual organization to make them productive on its own knowledge and its own processes. This week the supply side of the model market, from Thinking Machines’ Tinker to Fireworks’ specialized models, began building its revenue on exactly that capability. When the vendors start pricing a thesis, the thesis has stopped being a prediction.

None of this is a scoreboard of who deployed the most capital. The rounds matter as evidence for a thesis, and the sharpest formulation of that thesis came this week from the investors themselves: Union Square Ventures republished its old maxim as “Obliterate, Don’t Automate,” arguing that the biggest returns come not from automating existing work but from making the old process unnecessary. That is what robots on jobsites, models tuned on house knowledge and permission layers for machine colleagues have in common. They are not features added to work as we knew it. They are the early pieces of work rebuilt. Whoever waits for the finished picture will find that it was already there, and that they did not see it coming.

Gerhard Kürner is CEO of 506.ai, the European platform for Service-as-a-Software and agentic engineering, and author of Work After AI. Around 1,000 conversations with boards, owners, and PE funds across DACH and Europe.

Work After AI Weekly, by Gerhard Kürner. Every Friday I read the week’s signals from two sources, the investments of early-stage investors and the findings of research institutions, the same evidence base my annual report on AI and work is built on. Not because the investors themselves are the story, but because their investments are one of the clearest early signals we have for where work, organisations and enterprise software are actually heading.

Over the past three weeks this series watched companies hire an AI workforce, build a management structure over it, and hand it its first job with real professional liability. This week the digital colleague stops being a project and starts appearing in the systems that define what an organisation actually is. It gets an identity in the access directory, the same way a human hire does on day one. It starts changing who owns the business, because a professional whose entire back office is run by software no longer needs an employer. And the capacity it runs on is now being bought the way industrial firms buy factory capacity, years ahead and in billions.

The digital colleague now needs a badge, not just a login

Every organisation has a quiet system that decides who is allowed to do what, and until now it only had to answer for people. That assumption is breaking. Once dozens of AI actors work inside a company, each one needs an identity, a scope of permissions, and someone accountable for what it does with them, and managing that sprawl is becoming its own discipline rather than a checkbox in the IT department.

The market is already pricing this in. Oak, a company building what it calls an AI-native identity operating system, one control plane that manages humans, machines and AI agents as equal citizens of the directory, came out of stealth this week with a 60 million dollar seed round co-led by Accel, CRV and Greylock, unusually large for a seed and closed quietly back in late 2025. The banks are living the same shift in plainer language: Reuters reported this week that BNY now treats its digital employees as team members with their own logins and a human manager, and that other large banks are testing agents for core processes. When a bank gives a piece of software a login, a nickname and a supervisor, the identity system has become the org chart.

The professional keeps the license, the AI runs the office

Vertical AI spent the last two years selling efficiency to large employers. The more interesting move this week points the other way. When an AI operating system can run scheduling, billing, referrals and prior authorisations on its own, the licensed professional no longer needs the institution that used to provide all of that around them. The technology stops making employees more productive and starts making them owners.

Corner Health, which builds a network of primary care practices owned by the nurse practitioners who run them, raised 32.5 million dollars across seed and Series A this week, with Oak HC/FT leading the Series A. The company says ninety percent of its clinics operate without additional staff, a figure from its own announcement and worth reading as such, because the administrative work is carried by Cora, its AI-native practice operating system. The same pattern is arriving at institutional scale: Bunkerhill Health raised a 25 million dollar Series B led by Khosla Ventures this week for a platform that lets hospitals build their own AI actors instead of buying finished ones, with twenty-two agents already productive at one Texas health system.

This is the pattern the coming annual report calls the capability question. Work After AI, the yearly review of artificial intelligence and work that I will publish this autumn, argues that the decisive variable is not which model an organisation licenses but whether it builds the capability to make the technology productive inside its own walls. A one-clinic practice in Arizona and the Mayo Clinic are, on this evidence, now working on exactly that same problem.

Compute is bought like factory capacity now

A digital workforce has a means of production, and it is not the model. It is the inference capacity the model runs on, and the companies planning to field that workforce at scale have stopped treating it as a utility bill and started treating it as strategic procurement, contracted years ahead.

Reflection AI, backed by CRV since its seed round, signed a compute agreement worth more than 1 billion dollars with Nebius this week, running to 2029, stacked on top of an earlier capacity deal with SpaceX’s Colossus 2 worth up to 6.3 billion dollars, all to train open-weight frontier models and run its coding agent. Even the capital itself is retooling: Paradigm, long the technical flagship of crypto investing, closed a 1.2 billion dollar fourth fund the week before this window with a mandate that now spans AI and robotics alongside crypto. When specialist capital that committed redraws its own mandate, that says where the next decade of production capacity is being built.

The deals, at a glance

• Oak (Accel, CRV, Greylock, 60M seed, out of stealth): AI agents become identities to be managed like employees.

• Corner Health (Oak HC/FT, 32.5M seed plus Series A): the AI office turns licensed professionals into owners.

• Reflection AI (CRV, 1B+ compute deal with Nebius to 2029): inference capacity becomes strategic procurement.

• Bunkerhill Health (Khosla Ventures, 25M Series B): hospitals build their own AI actors in-house.

• Chai Discovery (Battery Ventures participating, 400M Series C): vertical AI reaches regulated pharma research.

• Paradigm Fund IV (1.2B new fund, week prior): specialist crypto capital retools toward AI and robotics.

Signals at the edge

Who writes the rules for the digital workforce? Within days of each other, two documents mapped the governance gap at the model layer from opposite directions. Google DeepMind chief Demis Hassabis proposed a FINRA-style, largely industry-funded standards body that would test frontier models before release and could coordinate a slowdown, with compliance eventually required for access to the US market regardless of a model’s origin. Days earlier the Future of Life Institute, an advocacy nonprofit with a declared regulatory agenda, published its summer index, in which an independent expert panel concluded that self-governance is no longer credible without external verification. No company scored above C+, the leaders have walked back their earlier pause commitments, and Europe’s flagship Mistral came last even as the EU leads on regulation, though Mistral objects that the scoring logic does not fit its development approach. Read together, the proposal and the report card make the same point: the rules of the field are being written elsewhere, and safety capability, like productive capability, is built inside the organisation rather than conferred by regulation.

Elsewhere, energy remains the binding constraint. The PJM capacity auction, covering the largest US grid, cleared at a record 16.4 billion dollars this week, with data centres accounting for roughly 6.3 billion of it. In defense, Helsing raised a 1.8 billion dollar Series E at an 18 billion dollar valuation, co-led by Lightspeed and General Catalyst, the largest European venture round on record. And on the labour side, Challenger, Gray and Christmas reported that AI remained the most cited reason for US job cuts in June, even as total announced cuts ran well below last year’s pace.

What this leaves on the table

Most of the firms in this week’s scan wrote no verifiable checks in the window, and that quiet is itself information: the money has moved down the stack, into identity, ownership structures and capacity contracts, the parts of the digital workforce that never appear in a product demo. None of this is a scoreboard of who invested how much. It is evidence for a single claim: the organisations taking this seriously are no longer asking what the technology can do. They are asking what has to be true inside their own walls, in their access systems, their ownership models and their procurement, for it to do that work reliably. Whoever waits for that question to answer itself will find, a few reporting cycles from now, that it was answered long ago, inside someone else’s walls.

Gerhard Kürner is CEO of 506.ai, the European platform for Service-as-a-Software and agentic engineering. More than 1,000 conversations over the last three years with boards, owners, and PE funds across DACH and Europe.

Two weeks ago companies started hiring an AI workforce. Last week that workforce got its first management structure, someone to certify it, someone to watch it, someone to keep it affordable. This week answers a harder question. What actually has to happen before a company hands a digital colleague something that matters, a task where a mistake is not an embarrassment but a liability.

Three deals this week trace that path in order. First a place to practice before anything is real. Then an engine big enough to run the practice at scale. And then, for the first time in this series, proof that the approach holds up in a profession where being wrong has always carried a price.

The practice room before the real job

No employer, human or otherwise, hands a new hire the keys on day one. There is a training period, a supervised stretch where mistakes are cheap and reversible, before anyone is trusted with something that cannot be undone. Companies building autonomous colleagues are discovering they need the same thing, and they are building it deliberately instead of hoping the model figures it out live in production.

You can already see this priced into the market. Wing VC led a 40 million dollar Series A into Bespoke Labs, with Mayfield Fund among the backers, for a company that builds the simulated environments and evaluation benchmarks agents are tested against before anyone lets them near a real workflow. The company is behind Terminal-Bench and OpenThoughts, benchmarks used across the industry to check whether an agent actually does what it claims before it gets production access. The pitch is not a smarter agent. It is a rehearsal space, and the fact that investors are funding the rehearsal space separately from the agent itself tells you how seriously the industry now takes the gap between demo and deployment.

The engine that cannot run out

A practice room only matters if the company can afford to run what graduates from it at scale, across every function, every day. That turns out to be a harder problem than picking the right model. It is a hardware problem, and this week’s biggest check went straight at it.

General Atlantic led the first close of a 1 billion dollar Series F into SambaNova Systems at an 11 billion dollar valuation, alongside Seligman Ventures, T. Rowe Price and Capital Group. SambaNova builds dedicated inference hardware, a direct alternative to renting Nvidia capacity, aimed at the fact that a company running a real digital workforce is not making one API call at a time anymore, it is running continuous inference across dozens of agents doing real work all day. The bet is not on which model wins. It is on the plumbing underneath, the same plumbing that decided last week’s story about whether an autonomous colleague is actually cheaper than the human it replaced.

The first regulated profession says yes

Everything above is preparation. This is the test case that matters, because it is the first deal in this series where the job on the other end carries real professional liability if it goes wrong.

Khosla Ventures led a 120 million dollar Series C into Norm AI at a 1.2 billion dollar valuation, with Coatue, Bain Capital Ventures, Blackstone, Craft Ventures and others participating. Norm AI builds AI agents that draft legal documents and handle compliance work under the supervision of human lawyers at its affiliated firm, Norm Law, which already serves clients with a combined 30 trillion dollars under management. Legal work has long been treated as the profession automation could not touch, because a mistake does not just cost time, it creates liability that lands on a licensed human. What changes here is not that the agent replaces the lawyer’s judgment. It is that the firm’s business model is shifting from billing by the hour to pricing by outcome, which only works if the agent doing the underlying drafting is reliable enough that the human supervisor is checking judgment calls, not fixing basic errors. That same structure, agent does the volume, licensed human owns the liability, outcome-based pricing replaces the billable hour, is the template every other regulated professional service is watching right now, audit and compliance work included.

The deals, at a glance

• General Atlantic → SambaNova Systems: 1B (Series F, first close). Inference hardware becomes the enterprise AI bottleneck.

• Khosla Ventures (lead), Coatue → Norm AI: 120M (Series C). A regulated profession goes agentic, with liability still human.

• Wing VC (lead), Mayfield Fund → Bespoke Labs: 40M (Series A). Practice environments before agents get real authority.

Signals at the edge

Three background trends are worth tracking even though they sit outside the work lens this week. Energy remains the structural constraint behind every deal above. A US heatwave this week strained the power grid supporting data centre growth, hyperscaler AI capital spending is on pace for roughly 750 billion dollars in 2026, and the Department of Energy expects data centres to draw as much as twelve percent of US electricity by 2028, up from four percent today, which means the inference capacity SambaNova and others are building is itself rationed by the grid. Defense tech funding has already passed 14.6 billion dollars in 2026, ahead of the entire prior record year, with Anduril alone raising a 5 billion dollar round at a 30.5 billion dollar valuation, but Fortune reported this week that investors in the category are now openly asking whether defense tech has become its own bubble stacked on top of the wider AI one. And the macro mood has not cooled. An internal US Treasury report is reportedly warning that the AI market shares structural risk with the dotcom bubble, a warning echoed this year by the Bank for International Settlements, even as AI companies keep pulling in a growing share of all venture capital.

What this leaves on the table

Put the three deals together and a clearer picture of what it actually takes to trust a digital colleague with something real starts to form. It has to practice somewhere safe first. It has to run on capacity that will not buckle the moment it scales past a pilot. And even then, in the one profession that put this to a real test this week, a licensed human still owns the outcome, just fewer of the hours that lead to it.

That is a more demanding bar than most companies are currently building toward. Plenty of AI rollouts skip the practice room and go straight to production, skip the capacity planning and hope the vendor’s infrastructure holds, and skip the question of who is actually liable when the agent gets something wrong. The capital in this week’s data is not a scoreboard of who wrote which check. It is an early map of what the companies actually getting this right are building first, the parts that do not show up in a product demo. Whoever keeps skipping that groundwork will find, a few reporting cycles from now, that it was there all along, just built by someone else first.

Gerhard Kürner is CEO of 506.ai, the European platform for Service-as-a-Software and agentic engineering. More than 1,000 conversations over the last three years with boards, owners, and PE funds across DACH and Europe.

Work After AI Weekly, by Gerhard Kürner. Every Friday I look at where the most sophisticated capital in the world is placing its bets, not because the investors themselves are the story, but because their bets are one of the clearest early signals we have for where work, organisations and enterprise software are actually heading. This week’s signal comes from the deals of the firms ranked eleven to twenty in the 2026 Strebulaev-Jackson Venture Ranking.

Here is where it starts. Companies have already begun hiring an AI workforce, software that finishes a task end to end instead of assisting a human through it. What this week’s signal shows is what happens next, and it has nothing to do with which firm wrote which check. It is a story about what happens inside a company the moment an autonomous colleague becomes a permanent hire. Three questions show up immediately. Can you trust it. Who is watching it. And can you actually afford to run it.

Read together, the deals of the week are not really about new digital colleagues. They are about the layer that makes the colleagues a company already hired trustworthy, supervised, and cheap enough to keep.

From can it do the job to can you trust it

For two years the competition among AI systems has been about capability, what a model can do that the last one could not. That race is quietly ending. Once an AI system runs as a colleague instead of a chat window, capability stops being the thing that decides who wins. Nobody hires a brilliant colleague they cannot trust in front of a client, and a model that occasionally makes something up is not a colleague, it is a liability with a nice interface. Reliability is becoming the actual product.

You can already see this priced into the market. Khosla Ventures backed Scaled Cognition’s new enterprise AI model, purpose-built to avoid hallucination and stay policy compliant while running live customer service work, with a 100 million dollar Series A. The company does not pitch a smarter model. It pitches a model a business can put in front of a customer without a human checking every output first, and that is a different sale entirely.

The digital teammate gets its own supervisor

Once a company runs autonomous colleagues across more than one function, a second problem shows up fast. Somebody has to see what they are doing, which systems they touch, and be able to stop them when something goes wrong. That supervisory function is turning into its own category of enterprise software, not a checkbox feature bolted onto the agent itself, and it is the clearest organisational signal of the week. The org chart question is no longer only who reports to whom among people. It is who inside the company is accountable for what the KI-Kollege did this morning, and what tooling gives them the visibility to answer that honestly.

Two early deals show the shape of that new layer. General Catalyst backed Tsuga, which builds observability for enterprise AI agents, essentially the monitoring layer DevOps built for software, now applied to a workforce that is itself software. Runlayer, with Khosla Ventures among its backers, raised money to give companies a way to track what their AI agents are doing across every system and pull them back when needed.

The economics of the autonomous colleague

None of the above matters if running the digital teammate costs more than the human it replaced. Whether the agentic company is actually cheaper than the human one it is copying depends almost entirely on inference cost, which is why the unit economics of the autonomous colleague are turning into their own investment category, quietly, underneath the more visible agent deals.

Two deals this week point straight at that plumbing. Kleiner Perkins backed Sail Research, which builds infrastructure that runs long-running AI agents more efficiently on the hardware companies already have, aimed at the fact that an autonomous agent can burn fifty to five hundred times the tokens of a normal chat session. General Catalyst also backed Together AI’s push to make open-source models a viable, cheaper alternative to renting every token from a single vendor. Put simply, the capital now flowing into that plumbing is the clearest evidence yet that inference cost, not model quality, is where the real margin fight will happen.

The deals, at a glance

• General Catalyst → Together AI, 800M (Series C): Open-model infrastructure for running agents

• Khosla Ventures → Scaled Cognition, 100M (Series A): Reliability becomes the product

• Kleiner Perkins → Sail Research, 80M (Series A): The cost of running an autonomous colleague

• General Catalyst → Tsuga, 35M (Series A): Observability for the machine workforce

• Khosla Ventures (participant) → Runlayer, 30M (Series A): Governance layer to watch and stop agents

Signals at the edge

Three background trends are worth tracking even though they sit outside the work lens this week. Defense tech funding has already passed 14.6 billion dollars in the first five months of 2026, more than the entire prior record year, and Fortune reported this week that several investors are now openly asking whether the category has become a bubble of its own, layered on top of the wider AI bubble debate. Energy remains the quiet constraint behind every deal above, with hyperscalers on pace to spend roughly 700 billion dollars on AI buildouts this year against a power shortfall Morgan Stanley estimates at close to 49 gigawatts by 2028, meaning the compute that autonomous colleagues need is itself rationed. And the macro mood around AI valuations has not cooled, DeepMind chief Demis Hassabis said publicly this year that early-stage AI startups with little traction are raising at unsustainable prices, a warning that keeps compounding as institutional capital keeps concentrating almost entirely in AI.

What this leaves on the table

Put the three shifts together and the shape of the agentic company keeps sharpening. It is not enough to hire the digital colleague. Someone has to certify it is reliable enough to trust, someone has to supervise what it does all day, and someone has to make sure it is actually cheaper to run than the person it replaced. None of that is a model problem. It is a management problem.

That is the part most companies still get wrong when they talk about their AI roadmap. They treat oversight, trust and cost as implementation details to sort out after the pilot succeeds. The capital in this week’s data treats them as the product itself, and that is the more useful way to read these deals: not as a scoreboard of who invested how much, but as an early map of the management layer every company will need to build for its own AI colleagues. Whoever keeps waiting to build that layer will find, a few reporting cycles from now, that the company sitting next to theirs already had a supervisor in place for its machines, and never noticed it being hired.

Gerhard Kürner is CEO of 506.ai, the European platform for Service-as-a-Software and agentic engineering. More than 1,000 conversations over the last three years with boards, owners, and PE funds across DACH and Europe.

Over the last three years I have had more than a thousand conversations with boards, owners and PE funds across DACH and Europe, mostly about the same underlying question. Everyone can see that AI is changing what a company needs to get work done. Almost nobody has a reliable way to see where that change is actually heading next, before it shows up in their own industry.

I started looking for an early signal, and I found one in an unlikely place. Venture capital is one of the few markets where the smartest people in the room have to place a real bet, with real money, months or years before the rest of the market catches on. When the firms at the very top of the field, ranked by the Strebulaev-Jackson Venture Ranking, start moving capital in the same direction at the same time, that is worth paying attention to. Not because the investors are the story. Because their bets are one of the earliest, clearest signals available for where work, organisations and enterprise software are actually heading.

That is what Work After AI Weekly is. Every Friday I go through the latest deals from ten of the top hundred venture firms in the world, rotating through the full list of a hundred over the course of the year, and I read them for one thing only: what they tell us about the future of work. Not which model is smartest. Not which round was the biggest. What it means for how you organise a company when part of the workforce is software.

Each issue distils that into three core trends and a short section on signals at the edge, the energy, defense and macro shifts that sit underneath everything else. No deal roster for its own sake. No hype. The venture data is the method, not the subject. The subject is always the same question. What is the smartest capital in the world telling us about the future of work, and what should you actually do about it before your competitors do.

The first issue goes out right after this note. It looks at what happens the moment an AI colleague becomes a permanent hire, and who has to manage it. It is not a prediction. It is a reading of decisions that have already been made, by people who had to be right or lose their investors’ money.

If you run a company, sit on a board, or simply want to see the shift before it reaches your desk, this is written for you. No jargon, no consultant framing, just the pattern as I see it, every week.

Subscribe here on Substack for the full issue every Friday, and follow me on LinkedIn for the shorter version of the same signal.

Whoever waits to pay attention to this will find, a few reporting cycles from now, that the shift was already visible in the data, and they never saw it coming.

Gerhard Kürner is CEO of 506.ai, the European platform for Service-as-a-Software and agentic engineering. More than 1,000 conversations over the last three years with boards, owners, and PE funds across DACH and Europe.