Neolab, unable to train models, has started reselling GPUs.
Neolab, an emerging AI lab, is currently the hottest kind of venture project — referring to those AI startups founded by star researchers that first raise massive funding and then pour it into model development.
Born under a halo, they outwardly all work on cutting-edge technology, but well…
Recently, an "absurd" business model has begun to surface, drawing even Musk over to watch:
Those Neolabs that locked in large batches of compute in advance with massive funding, even without training a good model, can still make a hefty profit byselling GPUs.
Regarding this business opportunity, the founder of AI cloud computing unicorn Modal said:
Next year, we may see a large wave of Neolabs transform into Neoclouds (AI compute cloud service providers).
Netizens have already figured things out for the Neolabs: in the best case, they achieve a technical breakthrough and become the next OpenAI; second best, they pivot to enterprise vertical Agents to survive; and in the worst case, if their models fail in training, they can sublease GPUs to pay dividends to shareholders as a floor.
This is truly mastering the game of cycle and risk hedging.
The compute scalpers have arrived
In large model entrepreneurship, compute is the biggestsunk cost, often taking up half or even more than 70% of a Neolab's total spending.
To prevent their clusters from suddenly losing supply mid-training, most labs, after completing their seed or Series A rounds, make signing 1-3 year guaranteed lease contracts with cloud providers or large data centers their first order of business.
On the other side, due to various factors, the compute spot market is currentlypersistently tight.
H200 market rental rates once climbed by about 160%, next-generation compute dark horse B200 rose 120%, and even H100 spot prices have maintained a premium of around 35%.
Thus, facing the huge spread between buying and selling prices, many labs whose research has stalled and cannot get models to work have started acting as compute scalpers.
Besideslaunching advanced models and selling tokens, they have found three ways to make money.
The first is to directly resell long-term compute lease contracts to brokers at a premium, earning risk-free middleman margins;
the second is to list spot GPU instances for sale, splitting up idle nodes and selling short-term compute to small teams desperate for compute by the day or week;
the third is to run open-source models in a wrapper for inference, simply giving up on pretraining and directly pointing their chips at mainstream open-source large models, offering token-billed API services.
In short, they can make money no matter what, crushing it.
Faced with the temptation of monetizing compute this way, an associate professor from Carnegie Mellon University also admitted on x:
AI genius entrepreneurship, and Silicon Valley VCs start doing the math
As the most money-magnetizing species in the venture world over the past two years, Neolabs all play the game in the same uniform way.
They are usually founded by star chief scientists from top big tech companies such as OpenAI, DeepMind, and Meta. They never engage in tedious application-layer development; upon launch they directly declare their aim to tackle AGI, and with extremely high talent density, they can leverage funding in the hundreds of millions as early as the PPT stage.
However, according to the latest report from Dealroom, Europe's authoritative venture capital data institution, the world's 29 most typical AI Neolabs have cumulatively raised a total of $23 billion (RMB 166 billion), yet only some of them have actually launched commercialized products available for external sale —The proportion is less than 35%.
That is to say,At least 19 laboratoriesThey either remain at a closed research stage, or have only released a technical whitepaper or a demo.
A typical example is Safe Superintelligence (SSI), co-founded by former OpenAI chief scientist Ilya Sutskever.
Relying on the unparalleled technical aura of its founding team, SSI saw its valuation pushed up to 32 billion US dollars even though it had been founded less than a year earlier and its team numbered fewer than a few dozen people. This speed would have been hard to imagine before '23.
However, from its very founding, SSI stated publicly that before achieving 'Safe Superintelligence (SSI)', the team would absolutely not develop any commercial products aimed at consumers or enterprises.
The sentiment is certainly noble.
But a Silicon Valley VC ran the numbers: today, a single ordinary full training cycle for a foundation modelburns at least 10 million US dollars in compute.
If a company wants to earn that money back solely by later selling APIs and tokens little by little, it must, after deducting electricity, bandwidth and depreciation, cumulatively sell 10 trillion-level tokens of effective API calls to the market.
Moreover, with the current top big tech companies dumping open-source models at subsidized or even free prices, labs without supporting enterprise workflows and proprietary data are unlikely to recoup their 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 thatNeolab is one of the most overhyped investment areas at the moment.
In his view, Google's early success was like "throwing a match into a bonfire," whereas some Neolabs today are trying to "light a match with a pile of bonfires."
As he sees it, injecting tens of billions of dollars at once into early frontier labsactually artificially raises the bar for their commercial success.
One More Thing
As for this Neolab operation of reselling compute, Musk also came over to watch and comment.
The point is: buying this many cards is useless, because the current compute shortage isn't about cards,it's about 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 falling far behind.
As of the 10th month of this year, 86 data center projects worldwide have been canceled or severely delayed.
Musk also predicted yesterday that the Neocloud industry will see a wave of large-scale consolidation and bankruptcies.
Neolab's road to transformation is probably going to be a hard one…