Three events in one week drew the same line: AI compute is being bought, sold, and financed like a physical asset. Groq raised $350 million to complete its pivot from chipmaker to neocloud operator, Etched’s valuation doubled to $21 billion after a trading firm tested and bought its inference systems, and Nvidia published the case for AI factories as an investable asset class. Together they describe a market where the accelerator is less a product than a holding.
Neocloud is the industry’s name for companies that buy accelerated computing in bulk and rent it out, operating Nvidia systems and increasingly their own silicon as utilities. The model collapses the older distinction between a chip company, a cloud provider, and a financial investor: the operator buys hardware, sells capacity, and finances the next purchase with the last contract.
Chips become a rental business
Groq’s path shows the turn most clearly. The company built its own LPU inference chips to compete with Nvidia, then lost its founder and top talent to Nvidia in a $20 billion licensing deal at the end of last year. Today it operates 13 data centers across North America, Europe, the Middle East, and Asia Pacific, running Nvidia systems for more than 6 million developers and enterprises. The new round, led by investment firm Disruptive with planned participation from Nvidia, values the company at $3.5 billion, lower than the $6.9 billion of last September; Groq frames the difference as a new valuation for the post-licensing company rather than a markdown. With the $650 million it raised in June, Groq has taken in about $1 billion in roughly three months, and it intends to scale from 54 megawatts of capacity to more than 200 megawatts by 2027.
The stated ambition is explicit. “We are building Groq into the world’s leading AI inference cloud,” said Alex Davis, chairman and CEO of Disruptive, which leads the round. “Inference will without a doubt become the largest and most critical layer of AI infrastructure.” Inference is the better rental business on paper: it is steady, repeatable demand rather than a training sprint, and it prices per token or per hour. Groq says the fresh funds support customers seeking medium and larger clusters of Nvidia accelerated computing for both training and inference.
The risks sit on the asset side of the ledger. Neoclouds carry heavy capital expenditures, debt, and hardware that depreciates quickly, and investors remain concerned about their ability to turn growth into free cash flow; Groq’s financials are still private. The company now sits inside Nvidia’s ecosystem rather than outside it, the same position as CoreWeave, Lambda, and Nebius, which run Nvidia GPUs while Nvidia invests billions in them. The chipmaker finances its own customers, because every megawatt they bring online is another megawatt of demand for its hardware.
Compute as a financial asset
Etched shows how the financial layer is thickening around inference hardware. The startup raised $700 million at a $21 billion valuation on Tuesday, led by Jane Street, after the trading firm tested and bought its systems; Jane Street said it is excited to have its own rack running in its datacenter. Etched was valued at $5 billion in December, raised a Series C at $10.3 billion in July, and has now doubled again in a month, up nearly $11 billion. The company sells full systems it calls frontier inference clusters, built around a prefill chip that operates at low voltage and a cluster-scale memory interconnect that lets many chips share a memory pool at low latency. A trading firm buying such systems is a different kind of customer than a startup renting GPUs: it is treating inference hardware as infrastructure with predictable output, the way it treats network capacity or a trading floor.
Nvidia’s own messaging has caught up to the theme. In an August 11 post, Jensen Huang argued that AI factories can be financed as productive infrastructure rather than built project by project, citing the $500 billion in third-party capital the company plans to mobilize with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR. “Compute is revenue,” he wrote. The argument leans on the property that makes an asset class work: a factory built on a widely deployed architecture is flexible and fungible, with a deep market of potential offtakers that protects residual value, and software improvements keep arriving for hardware that is already installed. Neon Control covered the financing platforms when they were announced, and the CME Group compute futures that launch October 5 will give the same capacity a public price.
What the shift asks of buyers
When compute becomes an asset class, the rental contract becomes the default and ownership becomes the alternative. For developers, the practical consequence is price discovery: futures, financing platforms, and competing neoclouds will make the cost of a GPU-hour more transparent and more variable, and comparing providers on real numbers matters more as the market fragments. For buyers who want control, the counterargument is the one this site has tracked for months: local models running on owned hardware avoid the rental market entirely, at the cost of capability and scale.
The next milestones are concrete. Groq targets more than 200 megawatts by 2027, Jane Street’s rack is already running Etched hardware, the first regulated GPU futures begin trading on October 5, and the Nvidia financing platforms start mobilizing capital. None of this settles whether inference will become the most profitable layer of AI infrastructure, as Groq’s new lead investor argues, or whether the neocloud business can convert its assets into free cash flow. It does settle that the answer will be measured in balance sheets as well as benchmarks, which is what happens when a technology becomes an asset class.
