No Product, No Problem: Investors Place Billion-Dollar Bets on AI 'Neolabs'

NextFin News - Ben Spector walked into investor meetings last fall with no pitch deck, no near-term plan to make money, and not even a concrete idea for a hit artificial-intelligence product. The Stanford University PhD student still got his funding. His startup is the purest example of a new category now sweeping venture capital: the "neolab," an AI research company that raises hundreds of millions - sometimes billions - of dollars before shipping a product, signing a customer, or even building a website.

Across 2025 and into 2026, private AI companies raised a record $225.8 billion globally, nearly double the prior year, and an increasing share of that capital is flowing to labs whose only collateral is a founding team and a promise of a future breakthrough. The central question for investors is no longer whether a company has traction; it is whether the people behind it might rewrite the next generation of AI. That is a bet on talent and optionality rather than revenue - and it is a bet that will be right exactly until it is catastrophically wrong.

The Neolab Playbook: Raise First, Build Later

The neolab model inverts the traditional venture-capital sequence. Instead of product, then customers, then scale, the order is now reputation, then capital, then - eventually - a product. The numbers are large enough to force the market to take the category seriously.

Thinking Machines Lab, founded by former OpenAI chief technology officer Mira Murati in February 2025, closed a $2 billion seed round in July 2025 at a $12 billion valuation, led by Andreessen Horowitz with Nvidia, AMD, Cisco, ServiceNow, Accel and Jane Street among the backers. By November, the company was in early talks to raise again at roughly $50 billion, more than quadrupling its value less than a year after launch. Its first product, Tinker - a fine-tuning API for language models - did not launch until October 2025, and even then it was offered free at the start, with usage-based pricing to follow.

Safe Superintelligence, founded by OpenAI co-founder Ilya Sutskever, has never released a product at all. Yet the company has raised approximately $7 billion to $8 billion across three rounds: $1 billion at a $5 billion valuation in September 2024, $2 billion at about $32 billion in April 2025, and a $5 billion round backed by Nvidia in July 2026. When Meta Platforms tried to acquire the startup during a hiring blitz, Sutskever rebuffed the approach and took over as chief executive himself.

The pattern extends well beyond the headline names. Eric Zelikman, a researcher who left Elon Musk's xAI, sought $1 billion at a valuation between $4 billion and $5 billion for Humans&, a lab building models designed to understand human emotions. Isara, a nine-month-old company founded by a former OpenAI safety researcher and an Oxford University student, raised $94 million at a $650 million valuation with OpenAI, Michael Ovitz and Stanley Druckenmiller among its investors; its early demo showed roughly 2,000 AI agents coordinating to forecast gold prices.

The exuberance is not confined to mega-rounds. Crunchbase data show that more than 41% of the $38.4 billion invested in global seed funding in 2025 went to AI-focused categories, up from 30% in 2024 - meaning the neolab mentality has reached the earliest stage of venture investing, where companies are often little more than a founder and a slide deck.

The macro backdrop explains how such deals became possible. CB Insights data show global private AI funding hit a record $225.8 billion in 2025, with mega-rounds of $100 million or more accounting for 79% of the year's funding. OpenAI, Anthropic and xAI alone absorbed a combined $86.3 billion, or 38% of total AI investment. By the first half of 2026, global venture investment in AI reached approximately $430 billion in just six months, exceeding the entire $254 billion invested across 2025. Deal count did not meaningfully grow; the records came from giant rounds, not broader activity.

Why Rational Capital Is Chasing Unproven Labs

The neolab boom is not random exuberance; it rests on a specific theory of where value sits in the AI stack. That theory has three pillars: scarce talent, scarce compute, and the commoditization of everything in between.

Frontier AI researchers capable of training state-of-the-art models are among the scarcest assets in technology. A handful of people moving between labs can shift the trajectory of an entire field, as the industry has watched repeatedly since the launch of ChatGPT. For an investor, backing a neolab is a scaled acqui-hire wrapped in optionality: if the team produces the next architectural breakthrough, the payoff is asymmetric; if it does not, the talent itself retains value.

"VCs are not underwriting today's models. They are assuming that scale, plus a breakthrough no one has seen, will manifest into market creation and capture."

That line from venture firm NEA captures the underwriting shift precisely. The capital is not paying for current capability; it is paying for the right to participate in a future one.

Compute reinforces the same dynamic. Training frontier models now requires clusters so expensive that funding itself becomes the moat. A lab without a billion-dollar balance sheet cannot compete, which means the rounds are self-justifying: only the well-funded can play, so being well-funded becomes proof of viability. Nvidia's involvement across Thinking Machines, Safe Superintelligence and the wider ecosystem is not incidental; the chipmaker is simultaneously the toll collector and a shareholder in the traffic. In March 2026, Nvidia and Thinking Machines announced a multiyear partnership to deploy at least one gigawatt of next-generation Vera Rubin systems, with deployment targeted for early 2027 - a scale of commitment that only the largest AI labs have approached, and one that came with an additional undisclosed Nvidia investment on top of its existing stake.

The third pillar is commoditization. Open-weight models are closing the capability gap with closed labs at a fraction of the compute the leaders spent to open it. Training data per active parameter has grown 3.1 times per year since 2022, dataset sizes are doubling roughly every six months, and estimates suggest training runs longer than nine months will become structurally inefficient around 2027. The marginal pre-training dollar is buying less capability than the dollar before it.

That dynamic has a second-order consequence the market is still absorbing. If general intelligence becomes a commodity produced by a few capital-intensive foundries, then the durable value migrates to what cannot be copied: domain-specific data that no one else can assemble, reinforcement-learning environments and verifiers tuned to high-stakes workflows, and the behavior layer where models learn to act inside a customer's operations. NEA points to work showing that a model pre-trained on a domain-biased corpus and then fine-tuned with reinforcement learning can outperform the same model without that targeted loop - a technical argument for why a small specialist lab can beat a generalist giant inside its niche.

The neolab thesis, stripped of its hype, is therefore a bet that the next wave of value will be built by small teams owning a narrow wedge rather than large teams chasing general capability. The funding is not a valuation of today's product; it is a prepayment for the scarce inputs - researchers and compute - that any product will require tomorrow.

The Exit That Makes the Math Work

For the neolab model to hold, there must be a buyer at the end of the road. The strongest evidence so far comes from Europe. In May 2026, SAP announced a definitive agreement to acquire Prior Labs, a Freiburg-based startup founded out of German university research that pioneered tabular foundation models - AI purpose-built for structured business data rather than language. SAP committed to invest more than €1 billion over four years, with reports putting the upfront payment near €700 million plus a €300 million earnout. The deal closed on July 17, 2026.

Prior Labs is the neolab bull's ideal case: an academic team that published a Nature paper, released an open-source tool, TabPFN, with more than three million downloads, and produced a billion-euro outcome in under two years without ever building a mass-market product. SAP is not buying revenue; it is buying a capability it could not develop internally fast enough.

"True intelligence requires moving beyond correlation to understand causation. Answering 'What will happen?' is useful, but answering why it will happen is transformative."

That line from the SAP announcement states the strategic logic in one sentence. General large language models are weak at structured, causal reasoning over business data; a specialist lab is worth more inside an enterprise software giant than it would be alone. NEA notes that top-decile software companies now trade at roughly 15 to 18 times their next-twelve-months revenue, against 3 to 6 times for the rest of the index - premium multiples that make strategic acquisitions of AI-native intellectual property economically rational even before the target has scaled commercially.

The agentic layer adds a second potential exit path. The market for AI agents is projected to grow from $7.84 billion in 2025 to $52.62 billion by 2030, a 46.3% compound annual growth rate, according to estimates cited alongside Isara's funding. If agent orchestration becomes the next platform layer, early infrastructure labs could be acquired by every major cloud and software vendor racing to embed multi-agent workflows - or go public themselves. That expectation is part of what allows a nine-month-old company to command a $650 million valuation with no revenue.

The Counter-Thesis: Dot-Com Economics With Higher Stakes

The strongest argument against the neolab boom is that it is the dot-com bubble replayed with more expensive hardware. In the late 1990s, companies with no products and no revenue proliferated because capital was cheap and fear of missing out was contagious. Most folded before earning a dime. The same fate awaits most neolabs, and the concentration of capital makes the fallout potentially larger.

NEA's own analysis undercuts the romantic view. Of the roughly $1.68 trillion in aggregate private valuation now sitting in AI research labs, 93.6% is concentrated in a single bucket of frontier generalists - OpenAI, Anthropic, xAI and Safe Superintelligence. Every other segment - biotech, robotics, scientific discovery, voice, edge - combines for barely 6%. The "neolab market," in other words, is not a diversified market at all; it is one concentrated basket of four names, and every dollar flowing into the category assumes the next model will deliver a step change that temporarily captures share in an endless tug of war.

There is also the human problem. Capital cannot buy cohesion. In January 2026, two co-founders and a founding engineer left Thinking Machines Lab to return to OpenAI. Fidji Simo, who leads OpenAI's applications business, confirmed the move had been planned for weeks, but the optics were damaging: a company valued at $12 billion losing some of its architects before it had proved anything. High optionality and ambiguous product roadmaps create churn, not commitment. The very feature that makes neolabs attractive - the freedom to pursue a breakthrough without the discipline of customers - is also what makes them fragile.

The bull case answers this by pointing to the barrier to entry. The dot-com era was defined by low capital requirements - anyone could launch a website. Frontier AI is the opposite: it is one of the most capital-intensive businesses ever created, which limits the field to teams that can clear the funding bar. The risk is not that everyone can do it; it is that even the winners may not be able to monetize it. That distinction matters, but it does not eliminate the risk - it concentrates it in fewer, larger positions, exactly where a mistake is most expensive.

What Comes Next: A Structural Shift With a Cyclical Top

The neolab phenomenon contains two forces that must be separated. The shift in what investors underwrite - from revenue traction to researcher talent plus compute optionality - is structural. The intensity of the funding frenzy itself is cyclical, and it will mean-revert when capital becomes expensive or when the first wave of labs fails to convert reputation into revenue.

Three scenarios frame the path ahead. In the base case, mega-round funding continues through the next 12 to 18 months as strategic acquirers keep paying premium multiples and compute supply expands; a handful of neolabs ship products that win enterprise contracts, while the rest are quietly absorbed in acqui-hires. In the upside case, one or more productless labs deliver a genuine architectural breakthrough - a new reasoning loop, a new modality, or a domain-native architecture - and the category is vindicated as the correct way to fund moonshot research. In the downside case, the first wave fails to monetize, follow-on capital evaporates, and valuations compress toward revenue reality; the $430 billion poured into AI in the first half of 2026 becomes the denominator of a painful write-down cycle.

In the short term, the momentum is likely to persist. With AI venture funding hitting $430 billion in the first half of 2026 alone and strategic acquirers paying premium multiples, the incentive to launch a lab and raise early remains powerful. More mega-rounds and more productless valuations are the base case for the next 12 months.

Over the medium term, a shakeout is likely. Labs that cannot ship models or products that win enterprise contracts will see their follow-on capital dry up, and the talent that justified their valuations will migrate back to better-funded homes - the Thinking Machines pattern, repeated across the category. LPs in funds that arrived late to the trend, and application-layer startups competing for the same talent and capital, are the most exposed.

In the long term, the winners are most likely to be specialist labs that own a domain-biased data corpus and run an integrated reinforcement-learning and verification loop inside a high-stakes workflow - the Prior Labs shape rather than the frontier-generalist shape. Compute sellers such as Nvidia and cloud providers benefit regardless of which labs win; strategic acquirers with balance sheets benefit from the option to buy capability rather than build it.

The falsifying signal is concrete: if, within 18 to 24 months, the productless neolabs have neither shipped models that win material enterprise contracts nor produced strategic exits at meaningful multiples - and if venture capital rotates back to revenue-multiple underwriting - then the "talent is the moat" thesis is wrong, and this cycle will be remembered as a costly detour rather than a regime change.

The neolab era is not a rejection of fundamentals; it is a relocation of them. Investors are no longer underwriting what a company has built. They are underwriting who is building it, and whether that team can reach the next breakthrough before the money runs out. In AI, that is a rational bet - right up until the breakthrough belongs to someone else.

Explore more exclusive insights at nextfin.ai.