The Silicon Industrialists - Part 1: AI's Gilded Age

AI was sold as software. But its economics look more like railroads, oil fields, and electrical grids.

The companies building frontier AI need data centres, specialised chips, power contracts, and hundreds of billions in capital before the revenue is fully there. But the same technology is also eating the pricing systems that were supposed to pay for it. Seats matter less when agents do the work. Implementation fees shrink when implementation becomes automated. Integration loses value when code becomes a disposable output of intent.

That collision creates the central bind of the AI economy: the costs are becoming industrial just as the revenue mechanisms are becoming unstable.

This article follows that bind through the infrastructure owners funding the labs that depend on them, the venture funds turning civilisational rhetoric into back-office automation, the startups that begin at the frontier and resolve into familiar enterprise tools, and the open-source projects whose independence cannot survive the cost structure that sustains them.

---

In the spring of 2026, Anthropic, OpenAI, Google, and Microsoft all made a rather strange announcement with three of them within ten days of each other in May. Instead of the usual tooting their own horns about a shiny new model, the breaking of a benchmark, or a product killing another 1,000 startups, we just got low-key press releases about deployment teams with engineers and embedded implementation teams inside client operations.

Anthropic backed a $1.5 billion services company with Blackstone and Goldman Sachs. OpenAI raised money for a deployment company with TPG and Bain Capital and acquired a consulting firm to staff it. Google began hiring forward-deployed engineers, with reporting that the plan involved hundreds of roles while Microsoft launched a joint practice with Accenture.

This is rather unusual if you take the public story of AI at face value. AI is supposed to remove the need for human implementation work. A small team with AI can do what a large team used to do. If any of that were fully true, the last thing these companies would need is thousands of people going to sit in offices and do the work by hand.

The surface explanation is that enterprises are messy. Dirty data, old workflows, brittle permissions, institutional knowledge that lives in places no API can reach. The official process says the customer workflow lives in Salesforce, but in reality it lives in a spreadsheet maintained by one exhausted operations manager who has been there for twelve years and is the only person who knows which numbers are real.

All this is true but it does not explain why the response from the most powerful AI companies in the world is not an innovative AI solution to the problem but to do exactly what Accenture and Deloitte and IBM have done for decades. Obviously, every technology wave had consultants and integration teams. But the issue here is that the companies promising to automate implementation are rebuilding implementation as the revenue model.

And the same pattern repeats across the entire AI economy. The infrastructure owners spend hundreds of billions building the substrate and then fund the labs that run on it, creating dependencies neither side can escape. The venture funds raise capital with the language of civilisational transformation and deploy it into automating the same back-office workflows that have existed for decades. The startups that promise to explore the frontier ship familiar or outright mundane products or get absorbed by the platforms that were already dominant. Even the open-source projects that are supposed to be the alternative end up at the same enterprise destinations, through different doors.

The answer requires understanding two forces that are operating on AI simultaneously, colliding with each other, and together pushing the technology ecosystem into ever more regressive structures. These are the industrial economics of AI, a cost structure that looks nothing like software and everything like 19th century heavy industry. And the progressive dissolution of the value-extracting mechanisms that are supposed to pay for it all.

The two forces

Force one: industrial economics

For roughly twenty-five years, software-as-a-service was the closest thing the technology industry had to a perpetual motion machine. You wrote the code once and you hosted it on rented infrastructure. You charged per seat, per month with every new customer adding revenue with almost no cost while the marginal cost of serving the next customer was close to zero. A mature SaaS company could reach gross margins of 75 to 85 percent. The product did not need to be manufactured, warehoused, or shipped. The capital requirements were modest.

AI was presented as the continuation of this, but supercharged. It would have the same high margins, low distribution cost, APIs and subscriptions with just a new logo attached.

But it turns out frontier AI does not work like that. Before a single customer arrives, the product requires an enormous amount of physical capacity that has to be built, powered, and paid for upfront. A frontier model is created by running vast computations across thousands of specialised chips for weeks or months at a time. These are clusters of GPUs and TPUs (graphics and tensor processing units originally designed for rendering video games and scientific simulations) repurposed for the mathematics that makes neural networks learn. A single training run for a frontier model can cost hundreds of millions of dollars in compute alone.

Then there is inference which is the cost of actually running the model once it exists. Every time a user sends a message or an agent executes a task, the model performs a computation. Unlike traditional software, where serving a request means looking something up in a database, serving an AI request means running the input through billions of parameters. The marginal cost of serving the next customer is considerable, and it scales with usage.

Both training and inference need chips and these chips need data centres. A modern AI data centre is not a room full of servers. It is an industrial facility spread over hundreds of thousands of square feet, consuming hundreds of megawatts of electricity, requiring dedicated cooling systems, backup power generation, and in many cases its own connection to the electrical grid.

The exact size of the AI infrastructure bill is hard to pin down because hyperscalers do not separate “AI infrastructure” cleanly from ordinary cloud expansion, data-centre construction, power procurement, and networking. But public estimates now put hyperscaler AI-related capital expenditure above $500 billion, while direct AI-attributable revenue remains much harder to isolate. One venture estimate put AI-related infrastructure allocation at roughly $560 billion across 2023–2024 against about $35 billion in direct AI revenue, a debated ratio of around sixteen to one.

The exact ratio may move, but the shape of the problem does not. The infrastructure bill is arriving far ahead of the revenue base. No software business has ever operated like this. This is the domain of utilities, oil exploration and submarine cable networks. Google’s own share, announced at its developer conference this month, is $180 to $190 billion. That is six times what it spent in 2022.

This is a construction boom, an energy boom, a financing boom, and a balance-sheet boom wearing a software label. The companies building AI are scaling like utilities, telecom operators and oil companies. They are industries where the physical infrastructure comes first and the customers come second. In reality, it is the capital requirements that shape every decision about what gets built and for whom.

Force two: value dissolution

At the same time that AI is demanding industrial-scale investment, it is dissolving the value structures that are supposed to pay for it. The pricing mechanisms, the per-seat subscriptions, the implementation fees, the integration charges that software companies have relied on for a generation, they are all gradually shrinking.

The better AI gets, the more legacy value it destroys. But it does not create equivalent new monetary value. This is the paradox that strains every AI business model.

For twenty-five years, software companies charged for the code, the interface, the workflow, the integration work that connected systems to each other. Implementation was scarce because it was hard and slow and required specialised knowledge. That scarcity is now collapsing. Code is starting to look less like a durable asset and more like a disposable output of intent. Open-source projects replicate in weeks what enterprise vendors charged millions to build. An AI assistant running on consumer hardware can do what entire software teams used to do.

By early 2026, HubSpot had fallen over 55 percent from its January 2025 levels. Atlassian had more than halved. Adobe and Salesforce were both down around 30 percent. Fund managers who had been long on software began exiting positions specifically because AI agent capabilities had crossed a threshold that changed the economics of seat-based pricing.

Per-seat pricing was a twenty-five-year artifact of the human bottleneck. One person, one seat, one subscription. When an AI agent can invoke a capability thousands of times in an hour, the seat is no longer the unit that matters. The companies that sell discrete tools (Figma, GitHub, Adobe, Synopsys) see usage multiply as agents automate tasks at scale. But the companies that sell coordination surfaces (Salesforce, Slack, Workday, ServiceNow) see seat economics hollow out because the human operating unit is no longer the scarce resource. ServiceNow already reports that 50 percent of its new business comes from non-seat-based pricing. Salesforce's CEO has acknowledged that the one-seat, one-unit model is dead. HubSpot's stock fell 19 percent after reporting a pricing shift, with its CEO naming Claude Code as a competitive threat on the earnings call.

AI makes things cheaper, faster and more accessible. The better it works, the harder it is to charge for the things it replaces and the mechanisms that used to capture value (subscriptions, seats, implementation fees, integration charges) stop working. The value is not disappearing. It is trying to migrate toward new forms (consumption-based, distributed, usage-metered) but the capital structure described here is not designed to let those forms develop. It is designed to recapture the value through the mechanisms that already exist.

The Collision: The double bind

These two forces operate simultaneously on everything in the AI economy.

Industrial economics pushes AI toward concentration. The compute is expensive, the facilities are enormous, and the capital requirements around them are staggering. Only a handful of companies can afford to build at this scale, and those companies become the landlords that everyone else depends on. This force pushes toward centralisation, dependency, and control.

Value dissolution pushes in what seems like the opposite direction. It makes things cheaper, opens layers, erodes the ability to charge for implementation. This force pushes toward commoditisation, and the collapse of existing business models.

Together they produce something worse than either force alone. Industrial economics gates who can build while value dissolution gates who can charge. And different actors experience them differently. A hyperscaler feels the first as opportunity and the second as irrelevant while an application-layer startup feels the first as dependency and the second as existential threat. The double bind appears when you need to do both.

AI is a groundbreaking technology caught in 19th century industrial economics, being forced through legacy capture mechanisms that the technology itself is dissolving. The industrial costs demand revenue while at the same time the value dissolution destroys the revenue models. And the result is regression to increasingly older methods of value capture because the double bind leaves no other path that the current capital structure can follow.

The obvious counter-argument is that this is simply what capital-intensive technology transitions look like. Railroads had circular financing, consolidation, and regulatory evasion. Telecom had infrastructure monopolies and captured standards. Electricity had decades of brutal economics before the grid became a public utility. In each case, the regression was the mechanism through which the technology got deployed. The difference in AI is that the technology being financed is simultaneously dissolving the revenue models of the customers who are supposed to pay for it. Railroads did not destroy the value of the cargo they carried.

---

The Infrastructure Ouroboros

A traditional software company could bootstrap or raise a modest venture round and rent what it needed from AWS or Azure. The infrastructure was someone else's problem and the landlord was interchangeable.

But the frontier AI labs cannot do this.

The industrial economics side: The capital required to train and serve a frontier model is so large that no lab can fund it independently. OpenAI, Anthropic, and Google DeepMind consume compute at a scale that makes them among the largest customers their infrastructure providers have. And the infrastructure providers (Amazon, Google, Microsoft) are not neutral landlords. They are technology companies with their own AI models, products, and competitive ambitions.

Microsoft invested billions in OpenAI and became its exclusive cloud provider. Amazon invested billions in Anthropic and committed to providing custom-designed chips. Google invested in Anthropic too, offering TPU capacity and cloud infrastructure. And in each case, the investment came with a compute commitment. The lab would run its training and inference on the investor's infrastructure, often locked in for years.

For the hyperscalers, every dollar the lab spent on compute flowed back as cloud revenue. The multi-year commitments became backlog, set as contracted future revenue that could be reported to investors and analysts. The backlog justified building more data centres which in turn justified more capital expenditure. The capital expenditure became the growth story that kept the stock price rising. And the equity stake in the lab meant that if the lab succeeded, the hyperscaler captured value both as investor and as infrastructure supplier.

The result is a structure where the supplier and the customer are financially entangled in ways that make the relationship almost impossible to unwind. The lab cannot leave without losing the infrastructure its models run on. The cloud provider cannot let the lab fail without writing down the investment and losing one of its largest compute customers. Both sides need the other to keep growing.

And the hyperscalers did not just invest in the labs and host their models. They even started building their own competing models. Google has Gemini, Amazon has its own Nova foundation models, Microsoft is developing in-house AI capabilities alongside its OpenAI partnership. The infrastructure provider is simultaneously the lab's investor, its landlord, its distribution channel, and in most cases its direct competitor. The competitive tension varies in intensity, but the structural entanglement does not. The frontier labs are training their models on hardware owned by companies that are using what they learn from hosting those workloads to build rival products.

Anthropic now has five compute providers: Amazon, Google, Microsoft, FluidStack, and SpaceX. The standard framing is "diversifying compute dependencies." The structural framing is different: five companies, four of which operate competing AI models, are simultaneously hosting the workloads of the model that is currently beating them. Diversifying your dependencies is not the same as not having them.

Anthropic's deal with SpaceX gives it all of the compute capacity at Colossus and a growing share of Colossus II in Memphis: more than 300 megawatts of capacity and over 220,000 Nvidia GPUs. This is the facility xAI built to train Grok. But SpaceX did not treat the facility as useful only for Grok. It entered the same business that Google, Microsoft, and Amazon were already in, leasing the infrastructure to the labs and capturing value from the substrate regardless of which model sits on top.

Reporting on SpaceX's S-1 shows the economics of the new play. The AI segment recorded a Q1 2026 operating loss of roughly $2.47 billion on $818 million of revenue. Anthropic's compute agreement, reported at $1.25 billion per month through May 2029, gives SpaceX an anchor customer and gives Anthropic badly needed capacity. The agreement covers SpaceX's Colossus and Colossus II data-centre clusters and can be terminated by either side with 90 days' notice. The facility was built for the model race, and when the model race could not pay for the facility, it resolved into an infrastructure transaction.

Microsoft and OpenAI are the prototype of how these entanglements resolve. On April 27, 2026, the two companies restructured their partnership. Microsoft no longer pays a revenue share to OpenAI, removing a drag on Azure OpenAI margins. Microsoft’s licence to OpenAI’s models became non-exclusive and now runs through 2032. OpenAI’s revenue-share payments to Microsoft continue through 2030 at the same percentage, now subject to a total cap; reporting has placed that rate at 20 percent and the cap at $38 billion. Even the AGI clause, the strange provision under which OpenAI's board could declare AGI and cut off Microsoft's access to the resulting systems, was neutralised into ordinary commercial terms.

At Alphabet, the Q1 FY2026 earnings release contained an even stranger signal. Operating income was $39.7 billion. Other income was $37.7 billion. Alphabet said the gain came primarily from unrealised gains on non-marketable equity securities. These are non-cash, they reverse if valuations decline, and they do not affect operating performance. But the structural signal is important here: a company whose core business is search, advertising, and cloud services now reports income that depends significantly on the valuation trajectory of the private AI companies it has invested in and hosts. The filing does not name the companies behind the gains, but financial press and analysts attributed much of the write-up to Alphabet's stakes in companies such as SpaceX and Anthropic.

The model race now shows up in places that do not look like product revenue but more and more like setups that resemble a hedge fund. The hyperscaler invests in the lab and the lab commits to cloud and compute. The lab’s valuation rises so the hyperscaler marks up the equity stake. The loop can improve reported earnings while the actual cash-flow dynamics of the AI buildout remain harder to see.

The value dissolution side: All of this infrastructure exists because AI dissolved software value mechanisms and created a gold rush. The current compute is being built because the collapse of software implementation scarcity triggered a scramble to find the next capture point, and every major technology company concluded that the answer was compute infrastructure. The $190 billion Google is spending is in reality a bet that whoever owns the physical substrate will capture value even as the software layer above it dissolves.

The double bind in action: The infrastructure owners do not need the model race to have a winner. They need the model race to run on their hardware. Under normal circumstances, costs this extreme would be the first thing the industry races to eliminate. You would expect a flood of research into efficiency, into smaller models that run on ordinary hardware, into architectures that need a fraction of the energy. And some of this innovation is happening, but that is not where the money flows. This is because every major participant in the loop is incentivised to make compute demand increase, not decrease. The hyperscalers need growing demand to justify the capex, cloud backlog and equity positions. Cheaper AI would be a genuine technical achievement but would also collapse the revenue structure that funds every company in the circle.

The industrial economics created the dependency. The value dissolution created the urgency. Together they produced a circular machine that needs appetite, not efficiency.

---

The Venture regression

In March 2025, Garry Tan claimed that a quarter of YC's Winter 2025 startups had used AI to generate 95 percent of their code. His conclusion was that founders no longer need 50 or 100 engineers. The same work could be produced with half a dozen people, maybe less.

But in early 2026, Marc Andreessen argued on the 20VC podcast that AI-driven labour displacement was “100 percent incorrect.” He described the layoff narrative as a farce, a cover story for correcting the overhiring of the zero-interest-rate years. In January 2026, a16z partner David George published a blog post arguing that the AI job apocalypse was a fantasy.

And the contradictions are getting louder. This month, Aaron Levie (founder of Box) predicted a consulting boom larger than the cloud transition. That is the same future the deployment-teams announcements are betting on, now coming out of the mouth of a SaaS CEO. In the same window, David Sacks argued on the All-In podcast that AI has dramatically lowered the cost of writing code but demand is rising so fast the net effect is more jobs, not fewer. He then described AI agents becoming cheap enough that enterprises could spin up alternatives to vertical SaaS, crushing the sales motion underneath those products.

Both can be true. But why are the people most committed to the 'AI-is-net-positive story' also openly describing the destruction of the categories AI replaces? What is going on here?

The value dissolution side: Tan is describing what AI actually does at the company level. A small team can now do what a large team used to do. Fewer engineers, fewer seats, less implementation work, less need for the old software machinery around big teams. That is genuine productivity but it is also economic compression. At the company level, this is fantastic. The startup operates at a fraction of the cost of a conventional competitor.

The industrial economics side: But the story that works at a company level causes devastation at the fund level. A giant fund does not want to hear that AI makes companies smaller, cheaper, and less dependent on vendors. It needs AI to mean more demand, more companies, more markets, and more exits.

A venture fund that raises $15 billion needs to deploy that capital. Management fees run at roughly 1.5 to 2 percent of committed capital, which is $225 to $300 million per year in fees before a single investment returns anything. The fund needs enormous rounds into enormous companies to justify the enormous capital base. And it needs a story big enough to raise the next fund.

The concentration is visible in the deal structure itself. In 2025, AI startups absorbed approximately 65 percent of venture deal value in the United States. And this capital is piling into a shrinking number of very large rounds. Meanwhile, PitchBook data shows that 15.9 percent of all US venture deals in 2025 were down rounds (a decade high) and AI/ML companies accounted for 29.3 percent of those down rounds. The startups that raised at peak valuations on frontier promises are hitting a funding wall when they return to the market with enterprise workflow products.

The double bind in action: Sequoia Capital declared long-horizon AI agents "functionally AGI." Andreessen Horowitz backed a startup whose launch materials ranked AI alongside fire, agriculture, and electricity. Benchmark told institutional investors that waiting on AI is "a path to extinction." The capital these firms deployed bought Hebbia, a search tool that helps investment bankers read SEC filings faster. Harvey, which automates contract compliance checks and legal citation formatting, at an $11 billion valuation. Viktor, which lives inside Slack and generates reports.

Some of these products may be genuinely useful, but they are not fire or AGI. They are not new categories of human activity. They are software tools that automate existing back-office work, the same work that has existed since the dawn of the modern corporation. They are things you could show to Jack Lemmon's insurance company clerk character from the 1960 movie "The Apartment" and he would immediately recognise them.

And automating that work may be genuinely valuable. The administrative overhead of the global economy is a multi-trillion-dollar drag on productivity. Boring automation has historically produced enormous structural change. Things like the shipping container restructured world trade without anyone calling it revolutionary. The problem is what the funding pattern crowds out.

And yes, venture capital does fund genuinely novel AI work with investments in things like humanoid robotics, robot foundation models, evolutionary architectures, biological computing. But these are not where the weight of the capital lands, they are the exceptions that make the pattern visible. The bulk of the money flows to back-office automation because back-office automation has something that genuinely novel AI ideas do not have. It is the easiest thing to fund.

They already have budgets, vendors, and managers looking for cost reductions. They already have procurement departments able to sign large contracts. And they already have incumbents who might buy the company later if the tool becomes dangerous enough.

In the first four months of 2026, global AI investment reached $233 billion already surpassing the cumulative full-year total for 2025. The overwhelming majority went to infrastructure platforms and horizontal foundation model companies. The genuinely novel categories, things like biological computing, neuromorphic hardware, federated learning, alternative architectures are rounding errors in the total. The Biological Computing Company, which cultivates living human neurons on electrode arrays and has demonstrated a 23x improvement in generative video model efficiency, raised $25 million in February 2026. In the same period, Harvey, Sequoia's legal document automation tool investment, raised $200 million at an $11 billion valuation.

These numbers do not describe an industry exploring the frontier. The venture machine is not even taking risks on new ways of doing business. AI compresses work and has the potential to dissolve existing corporate structures and replace them with entirely new, extremely compact ones. But the capital markets have found a simpler use for it: automating the workflows that already had budgets, vendors, and procurement departments ready to sign.

The industrial economics demand enormous capital. The value dissolution means the only legible deployment of that capital is wrapping AI around revenue that already exists. The technology may be new, but the business logic is old. Find an existing cost centre, insert "solution", capture the workflow and own the customer state. After you do all that you expand the contract and raise the next round.

But even this is not working any more, and the capital markets are already sorting it. OECD data shows venture funding for AI firms in IT infrastructure and hosting rising from $47.4 billion in 2024 to $109.3 billion in 2025. CB Insights data shows mega-rounds of $100 million or more accounting for 86 percent of all global venture dollars in Q1 2026, driven by enormous frontier AI rounds.

At the same time, the application layer is being repriced. B2B software multiples compressed sharply in early 2026.

The market is starting to reject the idea that every AI wrapper, workflow tool, or SaaS extension deserves frontier-lab economics. And as a result the money is retreating toward the layers where capture still works: compute, infrastructure, data centres, foundation models, and physical systems with capital barriers high enough to resist instant commoditisation.

---

From Frontier to Mundane

Not all AI funding goes to spreadsheet assistants and customer support chatbots. Some money does go to strange and in some cases genuinely exciting ideas. Biological neurons wired into silicon. Neuromorphic chips that escape the energy logic of GPU scaling. Photonic processors that compute with light. Federated learning systems that train models without centralising data. Neuro-symbolic architectures that combine language model fluency with deterministic reasoning. AI systems designed to operate autonomous laboratories.

Some of this work has real funding, serious teams, and genuine technical ambition that represent paths that could lead to forms of AI that look nothing like what the current market is building.

But even these exceptions frequently end up producing familiar market objects long before their original idea has had time to prove itself or fail on its own terms.

The value dissolution side: Alex Clayton of Meritech Capital described the core mechanism in June 2024. "It's almost a curse if you're an early AI company and you have revenue," he told the Wall Street Journal, "because then you get valued on the numbers, not just the story." As long as a company stays in research mode, its valuation is tied to an uncapped ceiling based on the perceived infinite potential of whatever frontier it claims to be approaching. The moment there is revenue, the valuation switches to standard multiples, and the numbers rarely justify what the story promised.

This creates a paradox. The company needs to show something legible to justify the next funding round. But the moment it shows something legible, it becomes that thing. And the thing it becomes is almost always familiar, because familiar is what the next round of investors know how to price.

In February 2025, Mira Murati founded Thinking Machines Lab. Murati had been CTO of OpenAI, one of the most visible leadership roles in the industry. The company entered carrying a promise of human-centred AI. Systems that would be more understandable, more customisable, more generally capable. The implicit market story was even larger. Here was an elite team led by a former OpenAI executive who might discover a new path beyond the existing model race.

In mid-2025, Andreessen Horowitz led a roughly $2 billion round valuing the company around $10 to $12 billion. The company had launched only months earlier, with no revenue and no product.

Then came the first product. In October 2025, Thinking Machines shipped Tinker, a fine-tuning platform for existing language models. Not a new model or a new architecture but developer utility built around the customisation of open-weight models, a category already served by dozens of platforms, several of them free.

By early 2026, the story had become even messier. Reporting described high-profile departures from the founding and technical team, including senior researchers returning to OpenAI and Meta. The company also struggled to secure a follow-on round at the much higher valuation it had reportedly sought.

As of May 2026, the company has restabilised under a new CTO, Soumith Chintala, co-creator of the PyTorch framework, expanded to roughly 140 employees, and is reportedly seeking a new funding round at a valuation of $50 to $60 billion, backed by major compute commitments from Nvidia and Google Cloud. The cycle may be reloading at higher stakes. But about a year and two billion dollars in, the only shipped product remains a fine-tuning platform for other people's models.

**The industrial economics side: ** Sometimes the company does not get pulled toward a boring product but torn apart when the infrastructure owner becomes the only plausible exit path.

Inflection AI raised a total of $1.525 billion to build Pi, a personal AI companion designed to be emotionally intelligent and fundamentally different from utility-first chatbots. Then Microsoft hired approximately 70 percent of the 70-person staff, including co-founders Mustafa Suleyman and Karén Simonyan. It paid $620 million for a non-exclusive licence to the models, and $30 million to waive legal claims related to the mass hiring. The company that raised $1.525 billion to build a fundamentally different kind of AI was hollowed out in a single transaction that its own regulators struggled to classify.

Adept AI raised $415 million to build general-purpose AI agents. Two of its co-founders had co-authored the original Transformer architecture paper, the foundational research behind today's generative AI. Amazon hired the cofounders and much of the technical team, leaving roughly a third of the company behind, and paid just $25 million for a non-exclusive technology licence. The remaining company continued under a reduced, restructured mission. The startup survived legally, but the centre of gravity had moved.

Startups being acquired by large technology companies is not new, it is the standard venture outcome. In previous technology cycles, acqui-hires happened to small teams that ran out of runway on modest funding. The buyer typically paid a premium, the deal was classified as an acquisition, and the outcome was pretty much understood by everyone involved.

But what is happening in AI is different. These companies raised enough capital to operate for years but they were absorbed before the public could learn whether the original mission had failed as a business or had simply become impossible to finance independently. Also the buyer was the company that was already providing the startup's infrastructure, had invested in it, and was even competing with it. And the transactions were designed from the start to avoid the legal and regulatory framework that exists to govern exactly this kind of consolidation. The Inflection AI deal was structured in a way that avoided ordinary Hart-Scott-Rodino premerger review. The FTC opened an investigation anyway while the UK CMA ruled it a de facto merger but cleared it. Its reasoning was that Inflection AI had not been a strong enough competitive constraint to matter.

That ruling raises its own question. Maybe these deals do not kill viable competitors but just revealed that the supposed competitors were already not viable. Maybe their frontier ambition could not survive contact with the industrial economics regardless of who came knocking. Whether these startups were genuinely pursuing their original missions or were never going to deliver on them is a question the public record cannot answer. The outcome is the same either way.

But the structure of the outcome is worth understanding. These deals are not simply conventional acquisitions adapted for AI. They are what happens when a full acquisition becomes both too expensive and too politically exposed.

For venture investors, it is not necessarily financial failure, in some cases they are made roughly whole. But it is a failure of the venture premise: the capital did not buy independence, category leadership, or a clean strategic exit but just enough time to create assets that the infrastructure owner could later extract. For the founders, the personal incentives often pointed elsewhere. Suleyman went on to run Microsoft AI and the remaining entity is not an independent company that chose a different direction. Once you extract 80 percent of the technical team and license the technology for a flat fee, what is left has fewer people to do the work, less leverage to raise capital, and little practical path back to the mission it was funded to pursue. The innovation goal is not deferred. It is structurally gone.

Conventional acquisitions still happen in AI. For example Databricks bought MosaicML for $1.3 billion in June 2023 and ServiceNow bought Moveworks for $2.85 billion. The hire-and-license structure appears when specific conditions converge: a startup valued too highly for a clean purchase, a buyer too dominant for a straightforward merger, and a product that has not yet become a business.

And the pattern is not confined to Inflection and Adept, though the cases vary in severity. Character.AI, Covariant, Windsurf, Hume AI, Contextual AI, and others all show versions of the same structure: talent migration plus non-exclusive licensing, often without a full corporate acquisition. Some leave active products behind or depleted shells while some later resolve into a conventional sale. But the recurring outcome is clear enough: many AI startups funded as independent frontier companies do not resolve as independent frontier companies but end up as licenses, teams, pivots, or remnants

The double bind in action: The hire-and-license playbook exists because a full acquisition is simultaneously too expensive and too exposed. The startup's inflated valuation made a clean purchase prohibitive. The buyer's market position as cloud provider, investor, and direct competitor made a straightforward merger politically dangerous. Nearly $2 billion in venture capital did not produce independent companies that explored the frontier but temporary structures that concentrated talent until the infrastructure owners could extract the parts that mattered. The venture machine and the infrastructure machine are just two stages of the same pipeline, and the output of that pipeline is not independent innovation but consolidation.

---

Capturing the open layer

The same double bind that pulls commercial AI toward short-term capture also acts on open source AI.

That sounds strange at first, because open projects appear to be the opposite of capture. They release free models, publish code, make tools available for developers, researchers, startups, and hobbyists. They lower barriers and create commons.

But openness does not sit outside the market just because the artefact is released publicly. An open model still has to be trained and local inference tools still have to be maintained. A developer framework still needs documentation, testing, security, and support and the economics leave very few exits. What follows maps five routes through which open AI projects end up serving the capture layer because the cost structure does not allow them to stay truly open.

And here is where the double bind bites hardest. Because open-source software had a version of this problem and solved it. Linux, Apache, PostgreSQL, Python, these are all projects that became critical infrastructure for the entire technology industry, built and maintained for years by communities of volunteers running code on commodity hardware. The marginal cost of keeping the project alive was low enough that independence was sustainable without corporate sponsors, venture capital or anyone's permission.

But AI has started to break this.

The industrial economics side: Open-source software could choose poverty and keep its independence. Open-source AI cannot because between the open artefact and the running system sits an infrastructure layer that does not obey software economics but the same industrial cost structure described in the earlier sections. The same physics that makes frontier labs dependent on hyperscalers also applies to open projects. Poverty means you stop training competitive models, and the moment you stop, you become irrelevant.

Traditional open source had the enterprise revenue pull with Red Hat and MongoDB as some of its most well-known examples. But open-source AI has that same pull, plus the infrastructure cost push which adds the question of who provides the physical substrate the work requires to exist at all. And the infrastructure is owned by the same small group of hyperscalers and chip companies that the rest of this article has been describing.

The value dissolution side: Traditional open source attracted corporate investment because companies could build value-capturing wrappers around the open artefact. The path was fund the project then sell the support, the hosting, the enterprise features and the managed deployment. The open code was free but the work around it was valuable enough to justify the investment. That was the Red Hat, MongoDB and Cloudera model. It worked because the wrapper held its price.

AI is dissolving those wrappers. The integration, customisation, deployment, and enterprise services that used to sustain the business model around open projects are themselves being commoditised by the same technology. If AI makes the wrapper worth less, there is less reason to fund the open project that sits inside it.

The double bind in action: The industrial economics make open AI impossible to sustain without outside capital. The value dissolution removes the traditional reason for providing it. What remains is strategic investment where capital arrives because the openness itself serves a capture strategy at a completely different layer. And the few open projects pursuing genuinely disruptive paths (smaller models, alternative architectures, ways to escape the industrial cost structure entirely) face an even starker version of the problem. Their success would undermine the compute dependency that the entire investment loop runs on. The entities with the capital to fund them are the same entities whose business models require the industrial economics to stay industrial. So these projects end up unfundable by the structure that exists.

1. Funding gravity

Mistral AI is the cleanest example. The company was founded in 2023 as the open European alternative to the closed American labs. Smaller models, efficient training, European sovereignty, open releases, less dependency on US hyperscalers. It was the kind of company that critics of the current AI economy could point to and say: the alternative exists.

Mistral needed compute to train larger models, infrastructure to serve them and capital to pay for both. It has now raised billions in equity, with Microsoft and Nvidia among its strategic relationships, and in March 2026 secured an $830 million debt facility to support its own AI data-centre buildout and Nvidia GPU procurement.

The company still matters as a European alternative. But the path to remaining an alternative now requires the same industrial apparatus: debt, chips, data centres, strategic investors, commercial licensing, and a product surface that gates the most capable systems behind paid access.

Stability AI, the company behind Stable Diffusion, followed a more dramatic version of the same arc. The open model created an enormous ecosystem. The company behind it burned through its funding, lost its CEO in March 2024, and nearly collapsed. That was until an $80 million rescue round, $100 million in forgiven debt, and $300 million in waived cloud liabilities stabilised the business under new leadership. The company survived. The open model ecosystem survived. But the company that emerged is an enterprise AI service running on AWS Bedrock, not the open generative commons it started as.

The same pattern holds across every open-weight project that began with independent intentions. Even the projects closest to traditional open-source independence follow the pattern. EleutherAI began as a volunteer community on Discord and needed CoreWeave's GPU clusters to train anything competitive. BLOOM mobilised over a thousand researchers across seventy countries and still required a three-million-euro government supercomputing grant to run the training. OLMo, led by the non-profit Allen Institute for AI, is funded by a $152 million joint grant from the US National Science Foundation and Nvidia, running on Nvidia's own Blackwell hardware. There is no competitive open-weight model that started independent and stayed independent. The project coordination is sometimes powered by the community but the compute never is.

2. Absorption

When an open project cannot sustain its own infrastructure costs and does not collapse, the most natural next step is absorption by an entity that already has the infrastructure.

MosaicML built tools for efficient open model training. Databricks acquired it for roughly $1.3 billion in June 2023. The open training layer became part of a large enterprise data platform. The promise changed from "anyone can train models efficiently" to "enterprises can build AI on Databricks."

Neural Magic built tools for sparse inference and model compression, including CPU-focused projects such as DeepSparse, SparseML, and SparseZoo. Red Hat acquired the company in January 2025. Within months, Neural Magic announced the end-of-life for those older CPU-focused tools and redirected attention toward vLLM and LLM Compressor inside Red Hat’s enterprise AI stack. A project associated with reducing dependence on specialised accelerator infrastructure ended up inside an enterprise infrastructure company, aligned with the model-serving layer that large customers were already adopting.

Nod.ai built open-source AI compiler tools. AMD acquired them in October 2023. The compiler was absorbed into AMD's Unified AI Software Stack and the open tools were redirected to optimise for AMD's own chips.

Determined AI built an open-source platform for managing deep learning training. HPE acquired them in June 2021. The training platform became a feature of HPE's GreenLake hybrid cloud services.

In each case, the open project existed to make some part of AI cheaper, more efficient, or less dependent on expensive proprietary infrastructure. The industrial economics made independence too costly. The value dissolution meant the open tool itself could not generate revenue. As a result, the projects ended up inside larger enterprise infrastructure companies, their technology redirected toward serving those companies' existing tech stacks.

3. The enterprise pivot

Not every open AI project trains models or runs data centres. Some are frameworks, libraries, and developer tools — the kind of project that in traditional open-source software could be maintained by a small team on modest resources. They do not need billions in compute. You would expect them to stay independent.

They do not. Not because the frameworks themselves are expensive to build, but because the production environments they live inside are. The AI ecosystem moves faster than any previous software category. Models change every few weeks. APIs break. New providers appear. Agent behaviour is nondeterministic. Every deployment creates security and data-governance questions that did not exist a year ago. The teams shipping AI into regulated, monitored, budgeted workflows need observability, evaluation, compliance, and production reliability — and those are the teams that can pay.

LangChain and LlamaIndex both began as open-source developer frameworks for building with large language models. Developers adopted them because they solved real problems: connecting models to tools, data, documents, workflows.

Then the paid layers appeared. Observability. Evaluation. Monitoring. Deployment. Enterprise support. Governance. Security. Compliance. The open framework remains. But the company sells the part that makes the framework safe and usable inside institutions. The website that used to say "build AI agents in Python" now says "enterprise AI platform" with a button that says "Book a demo."

This looks like the familiar open-core model. Red Hat gave away the code and sold the enterprise features. But traditional open-core often had a slower-moving technical core and a clearer separation between community edition and enterprise layer. AI collapses that separation because production reliability itself becomes the product. The enterprise layer is not additive. It is constitutive.

And the framework companies are not being pushed toward enterprise by their own costs. They are being pulled there by the double bind acting on everyone around them. The companies and developers building with these tools are themselves being pushed under the structural pressure of the double bind toward enterprise sales, compliance requirements, procurement processes. As the entire ecosystem moves toward production AI in regulated, budgeted workflows, the framework follows, because that is where the unresolved costs show up and the only audience that can pay. The code stays open. The hard part moved. The buyer changed. The roadmap followed.

4. Strategic openness

The three routes above describe what happens when open AI projects are pulled toward capture by the economics of sustaining themselves. But the dominant players are not just waiting for open projects to drift toward them. They are also building the open layer themselves.

Meta is the most visible case. In 2023, Meta began releasing Llama, a family of large language models whose trained weights are freely downloadable. Anyone — a developer, a startup, a researcher, a hobbyist — can download a Llama model, run it on their own hardware, modify it, and build products with it. The models come in different sizes, from small enough to run on a laptop to large enough to compete with the best proprietary models from OpenAI and Google. Training these models costs Meta hundreds of millions of dollars. It gives them away.

Meta's business is advertising. It does not sell AI models, charge for model access or compete with OpenAI's API and Anthropic's subscriptions. But still, the choice to spend insane amounts of money to train models just to make them open and free seems rather weird.

But by releasing Llama, Meta does two things. First, it ensures that Meta's own products are never dependent on a competitor's AI. Meta AI runs on Llama. Without it, Meta would have to license models from OpenAI or Google, paying a rival for capability inside its own platforms. The training cost replaces what would otherwise be a recurring licensing fee to a competitor.

Second, it undermines every company that does sell model access. If Llama is free and good enough, it puts downward pressure on what OpenAI, Anthropic, and Google can charge. A free alternative that anyone can run locally makes it harder to sustain premium pricing for proprietary APIs. Meta weakens the model sellers by making models abundant.

Meta's contribution to the commons is useful and developers genuinely benefit from free, powerful models. But the reason for Llama being out there has very little to do with the creation of an open-source AI ecosystem that would produce innovative solutions and more with its strategy against its competitors.

Nvidia follows the same logic at a different layer, and takes it further than anyone else.

Nvidia sells chips, specifically the GPUs that train and run AI models. It holds well above 80 percent of the data centre AI accelerator market. Revenue went from $17 billion in fiscal 2021 to $216 billion in fiscal 2026. Everything else Nvidia does in AI exists to make people buy more chips.

Beginning in 2025, Nvidia began releasing entire families of open source AI models across nine domains. Nemotron for agentic AI and enterprise workflows. Cosmos for physical AI and robotics simulation. GR00T for humanoid robots. Alpamayo for autonomous vehicles. Clara and BioNeMo for healthcare and biology. Earth-2 for climate science. Companies including Accenture, CrowdStrike, Deloitte, Oracle, Palantir, ServiceNow, and Siemens are already integrating Nemotron models into their operations. At CES 2026, Jensen Huang put it plainly: "We are now at the frontier of every single domain of AI models."

These models are more open than almost anything else in the industry. Nvidia publishes the model weights under a permissive licence that allows anyone to use, modify, distribute, and commercially deploy them without restriction. It publishes the training data (with over ten trillion tokens) and the training recipes. Unlike Meta's Llama, which restricts use by companies with more than 700 million users, Nvidia imposes no usage restrictions at all. The entire pipeline from data to trained model is available for anyone to inspect, reproduce, and build on.

But here is the catch. You can reproduce the entire model but you still need Nvidia GPUs to train and run it efficiently. The openness is total at the model layer while the capture is total at the hardware layer. Nvidia does not need license restrictions because the dependency is in the silicon.

Huang claimed that eighty percent of startups now build on open models. Nvidia is positioning itself as the company that leads that ecosystem across every major AI domain. Every model downloaded, every simulation run, every robot trained creates demand for Nvidia chips. Meta opens models in one domain to protect one business. Nvidia opens models across nine domains to drive demand for one product.

Meta and Nvidia are the most visible cases, but they are not the only ones. The same logic — open the layer you do not sell to drive demand for the layer you do — now describes nearly every major technology company producing open AI models.

Google releases Gemma, a family of lightweight open models, to funnel developers toward Google Cloud and its paid Gemini APIs. Microsoft releases Phi, a family of small open models, to drive Azure consumption and reduce its own dependence on OpenAI. Alibaba releases Qwen to drive adoption of Alibaba Cloud across Asia. The Abu Dhabi government funds the Falcon models as a sovereign capability play. DeepSeek, backed by a Chinese quantitative hedge fund, releases open models to power its own trading operations while compressing Western proprietary margins.

The industrial economics are what make this play exclusive to the largest companies. No startup can spend hundreds of millions to train a model and give it away as a strategic instrument. And the value dissolution is what makes the play rational: the more the model layer commoditises, the more it serves the interests of companies that capture value somewhere else entirely. Strategic openness is not available to small players and it is not designed to benefit them. It is a move only giants can make, and it works because the open layer accelerates the dissolution of everyone else's revenue model while leaving the giants' capture layers untouched.

5. The exception that proves the rule

Hugging Face seems like an outlier. Nobody absorbed it, it is not being pulled reluctantly toward enterprise or collapsing under infrastructure costs. It has investors with a cap table that reads like a roster of the companies described in this article: Salesforce, AWS, Google, Nvidia, Intel, AMD, Qualcomm, IBM, Coatue, Sequoia, while the founders retain roughly 25 to 30 percent voting control through super-voting shares. But the company's mission as the central hub of open AI has not visibly changed.

But the reason for this continuation in its commitment may be structural.

Hugging Face staying as the default hub where every model gets published and every researcher shares work, is itself the value proposition for the hyperscalers on its cap table. Every model downloaded from Hugging Face needs compute to run. Every Space needs inference. Every enterprise customer who discovers a model on the hub and wants to deploy it at scale ends up on AWS or Google Cloud or Azure.

The hyperscalers did not need to divert Hugging Face. The open mission is the demand generation engine. A locked-down Hugging Face would drive less adoption, less experimentation, less compute consumption. The openness is doing the work of customer acquisition for the infrastructure layer without anyone having to change a single governance decision.

Hugging Face is simultaneously the beneficiary of funding gravity, the institutional home that absorbs smaller open projects, and a case of strategic alignment where the genuine open mission serves investors' commercial interests without any diversion required. The project is faithfully executing its original purpose. The purpose itself has been absorbed into someone else's business model without changing a line of code.

Was all this inevitable?

This article has described two forces, the industrial economics and value dissolution, operating simultaneously on every actor in the AI economy. Where those forces meet, they produce a double bind that pushes the entire technology ecosystem toward regression. We have traced that double bind through four domains: infrastructure owners locked in circular dependencies with the labs they fund, a venture machine that deploys the language of transformation and the capital of automation, a portfolio system that converts frontier ambition into familiar enterprise products, and an open layer whose independence is structurally impossible to sustain outside the orbit of the companies it was supposed to be the alternative to.

The real problem with these incentives is not that the actors behind them are acting in bad faith but where the incentives lead when they coordinate. In 2025, four AI companies absorbed 65 percent of US venture deal value. A handful of infrastructure providers host every major frontier model. The open projects that were supposed to be the alternative run on the same infrastructure. And the most transformative technology in a generation is being used primarily to automate the invoice.

The question this raises is not whether any of it can be stopped. Much of it probably cannot. The infrastructure is already built, the funds are raised and the IPOs are next in line.

The question is whether it was inevitable. Whether the only possible path for a groundbreaking technology to evolve was by being bound to 19th century economics, existing capital structures, institutional infrastructure owners and enterprise procurement channels.

Whether anything like a different sequence is still possible, and what it would require, is the subject of Part 2.

---