NeoLab reselling GPUs becomes a new business model, as 29 labs raise $23 billion in funding

Born with a halo, on the surface they're all working on the most cutting-edge technology, but well...

Recently, an "outrageous" business model has begun to surface, and even Musk has come to watch:

Those Neolabs that used massive funding to lock in large batches of compute in advance can, even if they fail to train a good model, still make a hefty profit by selling GPUs.

Regarding this business opportunity, the founder of AI cloud computing unicorn Modal said:

Next year, we may see a large batch of Neolabs transform into Neoclouds (AI compute cloud service providers).

Netizens have already mapped out the fate of Neolabs: the best case is achieving a technological breakthrough and becoming the next OpenAI; the next best is pivoting to enterprise vertical Agents to survive by moving downmarket; the worst case is the model training collapsing, but they can still sublease GPUs and distribute dividends to shareholders as a fallback.

This is truly mastering the cycle and hedging risk.

Compute scalpers have arrived

For large model startups, compute is the largest sunk cost, often accounting for half or even over 70% of a Neolab's total expenses.

To prevent the cluster from suddenly being cut off mid-training, most labs, after completing seed or Series A funding, make their first move to sign 1-3 year minimum lease agreements with cloud providers or large data centers.

On the other side, affected by various factors, the spot compute market is currently persistently tight.

H200 market rental prices once climbed by about 160%, the new-generation compute dark horse B200 rose 120%, and even the spot price of H100 maintained a premium of around 35%.

So in the face of huge buy-sell price spreads, many labs whose research is stalled and can't get models to work have started acting as compute scalpers.

Beyond launching advanced models to sell tokens, they have found three ways to make money.

First, directly package long-term compute leases at a premium and resell them to brokers, earning a risk-free middleman spread;

Second, list spot GPU instances for sale, splitting idle nodes and selling short-term compute by the day or week to small teams desperately short of compute;

Third, wrap open-source models to run inference, simply abandoning pre-training and directly switching the chips to run mainstream open-source large models, providing token-billed API services.

In short, they can make money either way. Easy.

Faced with this temptation of monetizing compute, an associate professor from Carnegie Mellon University also admitted on X:

As the most fund-magnetizing species in venture capital over the past two years, Neolabs follow a uniform playbook.

They are typically founded by star chief scientists from top giants like OpenAI, DeepMind, and Meta, never bother with cumbersome application-layer development, and declare right at launch that they will tackle AGI. With extremely high talent density, they can often leverage hundreds of millions in funding at the PPT stage.

However, according to the latest report disclosed by Europe's authoritative venture capital data firm Dealroom, the world's 29 most typical AI Neolabs have cumulatively raised a total of $23 billion (166 billion RMB), and among them, those that have actually launched commercially sold products account for less than 35%.

In other words, at least 19 labs are either still in the closed research stage or have only released technical white papers or demos.

A fairly typical example is Safe Superintelligence (SSI), co-founded by former OpenAI chief scientist Ilya Sutskever.

Relying on the incomparable technical halo of the founding team, SSI, less than a year after its founding and with a team of fewer than a few dozen people, saw its valuation pushed up to $32 billion. This speed was hard to imagine before 2023.

However, from the very beginning, SSI stated externally that before achieving "Safe Superintelligence (SSI)," the team would absolutely not develop any consumer-grade or enterprise-grade commercial products.

Noble sentiments, to be sure.

But a Silicon Valley VC did the math: today, a single ordinary complete training cycle for a foundation model burns at least $10 million in compute.

If this money is to be slowly earned back purely through subsequently providing APIs and selling tokens, the company must, after deducting electricity, bandwidth, and depreciation, successfully sell a cumulative effective call volume of 10 trillion tokens to the market.

Moreover, against the backdrop of leading giants almost dumping open-source models in subsidized or even free form, it is not very realistic for labs without supporting enterprise workflows and monopoly data to recoup costs by selling APIs.

And according to Business Insider, Matt Murphy, a Silicon Valley veteran who participated in Google's early investment and led Menlo's investment in Anthropic, believes that Neolab is currently one of the most severely overhyped investment areas.

He believes that Google's early success was like "throwing a match into a bonfire," whereas some Neolabs today are trying to "use a pile of bonfires to light a match."

In his view, injecting tens of billions of dollars into early-stage frontier labs all at once has instead artificially raised the threshold for their commercial success.

One More Thing

Regarding this wave of Neolabs reselling compute, Musk also came to watch and comment.

The meaning is, buying so many cards is useless, because what's short in the current compute shortage isn't cards, it's electricity.

According to the latest Goldman Sachs research, global data center electricity demand is expected to grow 170% from 2025 to 2030, but grid construction, approvals, and infrastructure expansion are far from keeping up.

As of October this year, 86 data center projects worldwide have been canceled or severely delayed.

Musk also predicted yesterday that the Neocloud industry will face a wave of large-scale consolidation and bankruptcies.

The transformation path for Neolabs will probably be difficult to walk...