AMD just acquired a company whose chip is fast, inexpensive and a total mystery for lenders
Cheaper AI chips that are tailored to a specific model bring up new questions for creditors.
There is a chip that runs Meta's Llama 3.1 and nothing else. Not by configuration. The model is stamped into the silicon during manufacturing, and it will never run anything else for as long as it exists. It is remarkably fast, and by its maker's arithmetic, cheap to run. What nobody can tell you is what a bank would lend against it in three years.
The chip is called HC1, and on Aug. 6, AMD $(AMD)$ reached an agreement to purchase its maker, Toronto-based startup Taalas. With the deal, which is expected to close later this year, AMD is betting customers will accept single-model hardware if it is fast enough and cheap enough.
But every other chip in the data center can be pointed at whatever model comes next, which is the assumption underneath the current depreciation schedules the AI buildout is financed on. HC1 cannot. If its model falls out of favor, a lender left holding it has nothing to sell.
What the HC1 gives up
With reportedly more than 16,000 tokens per second per user, according to Taalas, the HC1 outperforms Nvidia's (NVDA) Blackwell generation hardware (roughly 350 tokens per second) and Cerebras's $(CBRS)$ wafer-scale chips at 2,000 tokens per second. To ensure the model would fit on the chip, Taalas stored its numbers at lower precision than a GPU uses, which costs some output quality. Even discounted, the speed is in a category of its own.
The cost figure carries the same qualifier. Taalas says running a million tokens on the HC1 would cost 0.75 cents, against 3.79 cents on an Nvidia GPU configured to serve as many users at once as possible. Both numbers are Taalas's own, and the cheaper one buys tokens from a compressed model.
Taalas makes the case that it has a lower cost of ownership than a GPU on a four-year refresh cycle, even if the Taalas chips have to be replaced every year, according to EE Times. That annual replacement is the trade. A GPU is more flexible than the HC1 could ever be: Once a model is outdated, the same GPU can spin up a different one, or do different work altogether. Taalas's chip cannot.
The money behind AI hardware purchases is increasingly borrowed. Nvidia is helping amass hundreds of billions of dollars of financing for AI infrastructure. CoreWeave (CRWV), which rents out GPU computing access to AI companies, has $2.6 billion of debt outstanding against its GPUs. Both deals rest on assumptions about what the hardware will be worth years from now.
When equipment is bought on credit, it usually doubles as collateral, or the asset a lender can seize if the loan defaults. That only works if the equipment has a resale market. A chip nobody else would want secures nothing, because a lender who seizes it has nothing to sell. The HC1 may be that kind of chip.
A chip's reusability used to be an engineering question. It has now become a financial one.
Nvidia is pitching adaptability as collateral
Four days after the Taalas announcement, Nvidia signed preliminary agreements with Apollo, Blackstone, Brookfield, BlackRock, Goldman Sachs and KKR to establish financing platforms intended to mobilize over $500 billion for AI infrastructure. CEO Jensen Huang said the company may guarantee up to $125 billion of it. No firm disclosed a pledge and the agreement still needs definitive contracts, so the pitch matters more than the number. Nvidia describes its hardware as an investment with the "lowest token cost, highest revenue and longest life." That is a list of collateral properties: flexibility across models and workloads, transferability between customers, and a useful life that its CUDA software platform keeps extending.
Colette Kress, Nvidia's CFO, made the same case on a November 2025 earnings call. Accelerators without Nvidia's software became obsolete within a few years as models changed, she said, while Nvidia's own A100 chips, shipped six years ago, are still fully in use.
That leaves the two biggest names in AI accelerators betting opposite ways on whether a chip needs to be flexible. Nvidia is raising half a trillion dollars on a pitch that its chips can be moved between models, workloads and customers. AMD has just bought a company whose entire cost advantage comes from giving that up.
Nobody can price the risk
Everything turns on how long the model stays in demand, and lenders cannot price the chip without an answer.
Paresh Kharya, VP of product at Taalas, told EE Times that he expects customers to stay with a model for a year or more. He works for Taalas, and the estimate is untested. It holds best for consumer applications, where consistent output matters more than leaps in quality, and worst for coding tools, where developers keep chasing the next best model. Faced with an unknown like that, a lender either declines or prices it in.
The buyer faces the same uncertainty. Companies spread equipment costs over the years they expect to use them, and the big cloud companies do not break out AI chips when they do. A shareholder cannot tell how long any particular chip is assumed to last. Amazon (AMZN) cut the assumed life of servers and network gear from six years to five, effective January 2025, citing the pace of advancement in AI and machine learning. That raised depreciation by $889 million and lowered net income by $677 million over the first nine months of that year. Meta $(META)$ went the other way, extending its estimate on most servers and network equipment to 5.5 years and cutting depreciation expense by $2.29 billion.
Two of the world's most sophisticated equipment buyers, depreciating similar gear, landed six months apart on how long that gear lasts. Model-specific silicon does not make the question easier. It ties residual value to the life of one model, and model lifespans are shorter and less predictable than server lifespans.
What a lender has already paid for
CoreWeave shows what that underwriting looks like when the hardware is reusable. The $2.6 billion it borrowed runs five years. Its customers sign three-year contracts, leaving two years of obligations with no contracted revenue behind them, a gap it has to close by re-signing those customers or finding new ones.
The lender advanced the money on the assumption that a GPU will still find a renter after the contract that paid for it expires, no matter what model is popular then. That assumption is the collateral. A chip built for a single model offers a weaker premise: a renter if its model is still wanted, nothing if it isn't. In a default, the lender holds an asset that lost its buyers at the moment it needed one.
What to monitor
For now, the gap in scale makes this a watch item, not an earnings story. AMD's data-center revenue more than doubled to $6.7 billion in the second quarter, with the Taalas deal still yet to close. The acquisition won't make an immediate financial splash once it's complete, seeing as Taalas is a 24-person company with $219 million raised since 2023 and one working chip as of February 2026.
Two things will tell you whether this hardware can become collateral. The first is whether cloud companies start reporting a separate lifespan for these chips, which would mean their accountants concluded the hardware is genuinely different. The second is whether financing terms show lenders assigning it value separately from the rest of the system.
That second point will be harder to read than the CoreWeave facility. AMD plans to sell Taalas technology inside systems alongside its Instinct GPUs, and a lender financing the whole rack reveals nothing about what it assigns to the model-specific part.
A Taalas chip could be cheaper to run than a GPU and still be more expensive to own, because the cost of the money is now part of the cost of running it. Taalas can price a million tokens to a fraction of a cent for as long as its model is wanted. Nobody has priced the years after. Until someone does, the industry's cheapest inference will be running on its least certain capital.
-Jurica Dujmovic
Comments