Radical Ventures AI Neolab Funding Hit $24 Billion, but Products Still Lag

Radical Ventures AI neolab funding reportedly reached $24 billion across the past two quarters, despite many recipients lacking established products, markets, or revenue. The total approaches five times the capital OpenAI and Anthropic raised before ChatGPT appeared, according to analysis highlighted by the Financial Times.

That comparison captures a sharp reversal in startup finance. OpenAI and Anthropic spent years developing research programs before their commercial momentum became obvious. Today, investors are placing billion-dollar bets on newly assembled teams before those teams establish product-market fit.

Radical Ventures calls these companies “NeoLabs,” meaning researcher-led startups pursuing foundational AI advances outside established frontier laboratories. The category includes Safe Superintelligence, Thinking Machines Lab, World Labs, Reflection AI, and dozens of more specialized research companies.

The wager is not simply that another chatbot will win market share. Investors are betting that small groups of recognized researchers can discover the architecture, training method, or learning system that replaces today’s dominant approach.

That thesis creates an uncomfortable question. Are venture firms financing the next OpenAI before its breakthrough, or paying frontier-company valuations for research projects whose commercial paths remain undefined?

Radical Ventures AI Neolab Funding Reached a New Extreme

The defining change is the amount of capital investors will commit before conventional startup evidence exists.

Radical Ventures identified more than 40 independent AI neolabs that have collectively raised over $40 billion since 2024. Its NeoLab market map places them across world models, reinforcement learning, continual learning, diffusion, and other research directions.

According to the Financial Times account, $24 billion arrived during the two most recent quarters alone. That pace matters more than the cumulative number because it suggests financing is accelerating rather than settling after an initial rush.

Radical Ventures offers a historical comparison. OpenAI and Anthropic together raised $5.3 billion between OpenAI’s founding in 2015 and ChatGPT’s November 2022 release. The new cohort attracted nearly five times that total within six months.

The comparison is imperfect. Compute costs have risen, investor familiarity with AI has changed, and later laboratories can recruit people with proven experience building frontier models. A dollar raised in 2026 also does not represent the same purchasing power or market context as a dollar raised several years earlier.

However, those qualifications do not erase the underlying shift. Startup funding once followed evidence that a company had discovered a market. In the AI neolab model, capital often arrives to fund the search for a discovery that might create a market later.

Safe Superintelligence illustrates the pattern. Ilya Sutskever, Daniel Gross, and Daniel Levy launched the company in June 2024 around one stated objective: building safe superintelligence. Within months, the laboratory had raised its first $1 billion.

SSI’s public site says its singular focus avoids “product cycles” and insulates its work from short-term commercial pressure. That research-first mission is unusually direct about postponing the standard obligations of a software startup.

Reports in 2025 said SSI subsequently raised another $2 billion at a $32 billion valuation. At that point, the company still had no public product roadmap. Investors were valuing the founders, the research objective, and the possibility of a future breakthrough.

Thinking Machines Lab followed another version of the same path. The company, founded by former OpenAI chief technology officer Mira Murati, closed a $2 billion seed round before releasing a product. Its early valuation reached $12 billion, according to contemporaneous reporting.

The size and timing of these rounds change what “seed investment” means. A seed round traditionally finances initial hiring, product development, and early market testing. A billion-dollar neolab round finances talent, compute, and time at something closer to institutional scale.

Radical Ventures says 13 North American and European model providers have completed rounds of at least $1 billion since the beginning of 2025. Eight belonged to its NeoLab category.

These are not isolated celebrity-founder deals. They form an emerging capital structure for AI research, one built around unusually large initial commitments and very long technical timelines.

Why Investors Fund Teams Before Products

AI neolab investors are underwriting scarce researchers and research theses, not current revenue.

That approach begins with the belief that a small number of researchers can materially change the direction of artificial intelligence. Their records become proxies for products that do not yet exist.

Sutskever helped lead the development of influential neural-network systems before co-founding OpenAI. Murati oversaw major product and research efforts at OpenAI. World Labs founder Fei-Fei Li has decades of recognized work in computer vision and spatial intelligence.

These backgrounds reduce one form of uncertainty. Investors can evaluate whether a founding team has previously contributed to frontier systems, recruited elite colleagues, and managed large computing programs.

They do not remove market uncertainty. An accomplished research team can pursue an idea that fails technically, arrives after a competitor, or proves too expensive for customers.

The investor argument rests partly on power-law returns. Most laboratories can fail if one successful company becomes a new platform with enormous economic value. A single winner might cover losses across the rest of a fund’s portfolio.

OpenAI and Anthropic serve as the existence proofs. Both emerged from concentrated groups of researchers, attracted substantial capital, and eventually built widely adopted products. Their trajectories make a new independent laboratory appear less implausible than it would have before ChatGPT.

Yet the comparison compresses years of difficult development into a clean success story. OpenAI explored robotics, reinforcement learning, language models, and other projects before ChatGPT created mass adoption. Anthropic developed its models and safety methods while building an enterprise-focused commercial operation.

New neolabs receive valuations influenced by what those companies became, not by what they looked like during their earliest years. Investors are effectively pulling expected future value forward.

Talent competition reinforces this behavior. A newly formed laboratory cannot hire recognized researchers using only a distant promise and ordinary startup compensation. It needs enough capital to offer competitive packages, secure computing capacity, and demonstrate that its work can continue through several training cycles.

Radical Ventures estimates that more than 40 NeoLab founders previously worked at Google DeepMind. Others have departed OpenAI, Meta, leading universities, and established AI startups.

DeepMind’s influence is especially important. Its teams contributed to AlphaGo, AlphaStar, AlphaFold, and other high-profile research programs. Former employees carry both technical knowledge and professional networks into new ventures.

This creates a circular funding mechanism. Famous researchers attract capital, the capital helps recruit more famous researchers, and the expanded team supports a higher valuation. Product evidence can remain secondary during the cycle’s early stages.

Investors also fear missing a category-defining company. Once a laboratory secures a renowned team and a large computing commitment, joining later can become much more expensive.

That fear favors rapid decisions. Waiting for revenue might produce better evidence, but it can also eliminate access to the investment or raise the entry valuation dramatically.

The result resembles financing for pharmaceutical research more than ordinary software. Investors fund scientific programs through long periods without commercial revenue because a successful discovery can support a valuable product portfolio.

AI differs in one crucial respect. Pharmaceutical development has structured clinical stages, regulatory milestones, and clearer methods for comparing candidate treatments. Frontier AI research lacks an equivalent validation framework.

Benchmarks offer partial evidence, but they can be optimized, contaminated, or overtaken quickly. Technical demonstrations reveal capability without proving that customers will pay enough to cover development and inference costs.

That gap explains why founder reputation carries so much financial weight. When objective milestones remain uncertain, prior achievements become an unusually important signal.

The Bet Is New Architectures Against the Current Frontier

The primary contest is between alternative research paradigms and the scaled transformer systems already commercialized by incumbents.

Most neolabs are not promising a slightly improved version of an existing assistant. They argue, explicitly or implicitly, that current models face limitations which require a different technical path.

World-model researchers believe systems need internal representations of physical environments, not only statistical patterns learned from text. Such models attempt to predict how a scene or environment changes after an action.

World Labs is pursuing spatial intelligence, while other companies are applying related ideas to robotics, video, and simulation. Their opportunity depends on current multimodal models reaching a ceiling that world-based representations can surpass.

Continual-learning companies focus on another limitation. Most deployed foundation models remain largely fixed after training. Developers can add retrieved information or fine-tune them, but the underlying model does not continuously update itself through everyday experience.

A successful continual-learning system would adapt while preserving earlier knowledge. That combination remains difficult because new learning can damage previously acquired capabilities, a problem often called catastrophic forgetting.

Reinforcement-learning laboratories make a different argument. They contend that models trained mainly to predict human-generated data cannot reliably move beyond existing knowledge.

Their systems learn from feedback inside environments where results can be checked. Winning a game, proving an equation, or producing a working program provides a stronger signal than imitating text alone.

Other groups are exploring diffusion-based language models. Traditional autoregressive models generate one token after another. Diffusion approaches can refine larger blocks in parallel, potentially changing the tradeoff between speed, cost, and output quality.

Energy-based models take yet another route. Rather than generating a sequence directly, they score candidate configurations against constraints. Advocates believe this process can improve planning and verifiable reasoning.

These approaches are technically distinct, but their investment stories share the same structure. Each laboratory identifies a limitation in today’s frontier models and proposes a method that might become essential after scaling delivers diminishing returns.

OpenAI, Anthropic, Google DeepMind, and Meta are not standing still while these startups conduct research. The incumbents can explore many of the same methods internally, supported by existing infrastructure, distribution, and customer relationships.

That is the central pressure on the neolab thesis. A startup needs more than a valid research idea. It must develop that idea faster than an incumbent can reproduce, acquire, or supersede it.

Radical Ventures lists this moving target among the category’s key risks. Coding laboratories, for example, can spend years developing specialized systems only to encounter rapidly improving products such as Claude Code or Codex.

Open-weight models add another challenge. If capable systems become openly available soon after a laboratory completes an expensive training run, technical differentiation can lose commercial value before the company builds distribution.

Research originality also does not guarantee a durable business. A laboratory might publish an influential technique that larger companies incorporate without paying the original developer.

The best outcome requires three connected achievements. The team must make a real technical advance, translate it into a product, and create a commercial advantage that competitors cannot quickly erase.

The first step attracts attention and financing. The second and third determine whether investors funded a company or an unusually expensive research program.

What the $24 Billion Figure Does Not Show

Headline funding totals can overstate how much independent risk investors have accepted and how much usable capital laboratories control.

Private financing announcements often combine several forms of commitment. A reported round can include cash, cloud credits, future tranches, strategic investments, or financing linked to specific conditions.

The headline amount therefore does not always equal unrestricted money deposited immediately into a company’s bank account. Without complete transaction documents, outside readers cannot reliably compare every round on identical terms.

Valuations can create another distortion. A small portion of a company may be sold at a high price, establishing a headline valuation for the entire business. That number does not show whether the company could sell a much larger stake at the same price.

Strategic investors also have interests beyond financial returns. Chipmakers benefit when laboratories purchase computing hardware. Cloud providers benefit when startups commit to long-term infrastructure use.

Nvidia has become one of the most active investors across the model-provider market, according to Radical Ventures. That position reflects the close connection between startup financing and compute demand.

Radical estimates that training and inference usually consume well over half of a neolab’s spending. If an investor supplies chips, cloud capacity, or financing tied to infrastructure, part of the transaction supports the investor’s own commercial ecosystem.

This does not make the investment artificial. Laboratories genuinely need compute, and strategic partnerships can reduce the risk of unavailable capacity. However, it complicates any simple interpretation of funding as independent confidence in future product revenue.

Compute commitments can also become liabilities. A laboratory may secure years of capacity before knowing which architecture will work or how much demand a future product will generate.

Training a frontier model produces a technical asset, but each new generation can require another large investment. A company that misses one important cycle can face pressure to raise again at unfavorable terms.

SSI’s original financing showed how quickly this tension can appear. Axios noted that even a billion-dollar cash round might not last long if the company pursued superintelligence without interim revenue. Its funding analysis identified repeated capital needs as a central risk.

The absence of revenue is not automatically evidence of failure. Research-oriented companies can rationally delay commercialization while protecting a difficult technical program.

The problem is that investors must distinguish productive patience from an indefinite absence of validation. Private reporting gives outsiders limited information about training results, spending rates, or internal milestones.

Laboratories can remain in stealth while raising at higher valuations. Each new round then appears to validate the last, even when the strongest signal is another investor’s willingness to participate.

The model becomes fragile if capital availability weakens. A company with paying customers can adjust spending, improve margins, or focus on its strongest product. A pre-revenue laboratory has fewer options because cutting compute or losing researchers can undermine its main asset.

Talent concentration introduces similar risk. A laboratory’s valuation may depend heavily on several recognized founders or senior researchers. If they leave, the company cannot replace their reputations as easily as a software business replaces a sales executive.

Thinking Machines encountered an early example when co-founder Andrew Tulloch returned to Meta after its major seed financing. The departure did not determine the company’s fate, but it showed how quickly a talent-based investment thesis can be tested.

Traditional startup diligence asks whether customers repeatedly choose a product. NeoLab diligence often asks whether a team can remain together long enough to discover something defensible.

Those questions demand different evidence. Funding totals answer neither one.

AI Neolab Funding Puts Incumbents and Smaller Startups Under Pressure

The money intensifies competition for researchers and compute while raising the minimum cost of pursuing frontier AI independently.

OpenAI, Anthropic, Google DeepMind, and Meta face a direct retention problem. Their strongest researchers now know that departure can unlock enormous funding and considerable control over a new research agenda.

Large laboratories can respond with higher compensation, greater autonomy, or internal teams organized around speculative projects. Meta’s aggressive recruitment campaigns showed how quickly researcher compensation can become a strategic expense.

Incumbents retain important advantages. They provide vast computing systems, experienced operations teams, access to proprietary data, and immediate paths from research to millions of users.

However, scale can also create frustration. Product schedules, safety reviews, enterprise commitments, and organizational layers can constrain researchers who want to pursue a narrower technical thesis.

NeoLabs sell an alternative. A small team can choose one long-term problem, avoid supporting mature products, and organize the entire company around its research.

That proposition pressures independent startups too. A conventional AI company seeking several million dollars to build a useful product now competes for attention with laboratories raising billions around more dramatic technical ambitions.

The two groups often pursue different markets, but they draw from overlapping pools of engineers, investors, and computing resources. High compensation at well-funded labs can make hiring much harder for product-focused companies.

Cloud and chip capacity create another divide. Radical Ventures argues that guaranteed, multi-year compute access has become more restrictive than capital itself.

A team can raise money and still struggle to obtain enough advanced accelerators at the right time. Delays can disrupt training schedules while better-connected rivals move forward.

The financing wave also affects enterprise buyers. More laboratories promise specialized models, alternative architectures, or lower-cost inference. Buyers gain options, but many providers lack proven operational histories.

A technical benchmark does not answer whether a vendor can support production workloads, protect data, meet service commitments, or survive several years. Procurement teams must evaluate the financial endurance of suppliers alongside model capability.

Developers face a related problem. Building deeply on a new model can create switching costs even when its application programming interface appears compatible with alternatives.

A neolab that changes direction, runs out of funding, or sells its team to an incumbent can leave customers without the product roadmap they expected.

Acquisition is not a guaranteed safety net. Radical Ventures notes that the pool of natural buyers is limited because major technology companies already maintain close relationships with established AI laboratories.

Regulators have also examined arrangements that transfer teams and technology without purchasing an entire company. These so-called reverse acquihires can give founders and employees an exit while leaving outside investors with uncertain recoveries.

Microsoft’s arrangement with Inflection AI demonstrated why such structures attract scrutiny. The transaction moved senior personnel and licensed technology without following the simplest form of a conventional acquisition.

For a successful neolab, pressure on incumbents can become productive. A credible alternative can force faster releases, better research conditions, or more competitive pricing.

For unsuccessful laboratories, the same pressure can produce expensive duplication. Multiple teams may spend heavily training similar systems before learning that none has a defensible advantage.

The $24 billion wave therefore increases both experimentation and waste. Venture portfolios can tolerate that combination, but infrastructure providers, employees, and enterprise customers experience the consequences differently.

Three Signals Will Test the AI Neolab Thesis

Products, independent technical evidence, and repeatable revenue will determine whether the financing wave created new frontier companies.

The first signal is product delivery. Laboratories that raised large early rounds must turn research programs into systems outside users can examine.

A release does not need to resemble ChatGPT. It can be a model, developer platform, robotics system, scientific tool, or infrastructure product. It must still solve a defined problem for users beyond the founding research team.

Thinking Machines offered an early test when it began releasing tools after its record seed round. Its company updates provide a public record against which investors and developers can compare earlier expectations.

The crucial question is not whether a laboratory launches something. Companies can package existing methods into respectable demonstrations. The question is whether the product shows an advantage connected to the laboratory’s original research thesis.

The second signal is independent technical validation. Neolabs frequently make claims about new learning methods, architectures, or paths toward general intelligence.

Those claims become more credible when outside researchers can reproduce results, customers can test performance, or detailed evaluations show advantages across meaningful tasks.

A benchmark lead that disappears after several weeks offers limited support for a multibillion-dollar valuation. A method that lowers training costs, improves reliability, or enables previously impractical applications provides stronger evidence.

Investors should also watch whether incumbents adopt similar methods. Adoption can validate the research direction while weakening the startup’s commercial moat.

The third signal is revenue quality. A laboratory does not need immediate profitability, but recurring customer demand separates an enduring company from a financed experiment.

Revenue should come from customers choosing the product, not mainly from related strategic arrangements or one-time research partnerships. Growth also needs to remain meaningful after accounting for inference and infrastructure costs.

OpenAI’s trajectory illustrates why distribution matters. Its 2025 financing announcement tied new capital to expanded infrastructure and products serving hundreds of millions of weekly users. The company’s funding update connected research spending to an already visible adoption engine.

Most new laboratories have not established that loop. They are raising the capital required to discover whether one exists.

The next several quarters should expose meaningful differences across the category. Some teams will release credible systems. Others will remain private while seeking additional funding. A few will lose talent, change direction, or pursue strategic exits.

The $24 billion total is therefore a starting point, not a scorecard. It measures investor willingness to finance alternative paths beyond today’s frontier companies. It does not measure how many of those paths work.

Developers and enterprise buyers should watch whether each laboratory’s product reflects its promised technical distinction. Investors should look beyond the next valuation and ask whether evidence is accumulating faster than spending.

Radical Ventures AI neolab funding has made independent frontier research possible at unprecedented scale. The harder task begins after the round closes: converting concentrated talent, costly compute, and scientific ambition into products that people repeatedly choose.

Which laboratories can show that conversion before their next major financing? That answer will reveal whether the NeoLab boom widened the AI frontier or merely raised the price of searching for it.