David López Mateos @SenorScience

Funny detail in here: the oldest chip, the one with the biggest install base, had the biggest rental jump this year. So much for old silicon being the depreciating trade. People miss it because there's no public price for any of this, rental or resale. You only catch it by tracking where it actually clears. Daily.

Compute Desk @ComputeDesk

a week ago

Daily GPU compute rates live from @Architect_Fi, ahead of imminent Compute Futures on AI Exchange. Tickers powered by @ComputeDesk. Reach out for hypergranular intraday feeds.

0 5 17 4K 3

Worth 99 seconds of anyone's time. Compute futures stopped being a thought experiment and turned into a live product faster than almost anyone expected. Good to see it explained plainly.

Jacob Robinson @JacobRobinsonJD

2 weeks ago

Compute futures are the most novel product I've heard of in the past month. What is a "compute future"? @BrettHarrison explains exactly how it works in 99 seconds (he and his team at @Architect_Fi pioneered them).

1 6 25 8K 7

Resale and yield run on different clocks. Resale tracks the secondary market for the chip. Yield tracks the rental market for the hour it produces, and that one is climbing on scarcity: spot discounts on H100 have been narrowing as providers reprice reclaim risk. A chip can be worth less while the compute it produces rents for more. There's a trade in the gap between the two.

Edge inference is here. The market for it isn't. Comcast, Charter, Akamai, Cloudflare, Nscale: real GPUs in real metro DCs. But the pricing layer to clear them doesn't exist. What programmatic ads solved, edge compute hasn't: computedesk.substack.com/p/edge-inferen…

The demand side is not the question: it is there. Whether these take off comes down to the settlement index: does it track the price an operator actually pays. Too broad and the hedge carries basis risk against an average nobody touches. Too narrow and it is too thin to settle. The index is the hard part, and it is the part we just built.

"The suitability of the Hedge depends on the index." @Architect_Fi compute futures powered by @ComputeDesk indexes.

Brett Harrison @BrettHarrison

3 weeks ago

The typical neocloud operator we speak to says some version of: “I’ve sold out 12 months of capacity on my H200 cluster, but I don’t know what compute prices will be in months 13-24.” Futures solve this, but the suitability of the hedge depends on the index. More below ↓

8 15 80 19K 31

The months 13-24 line is exactly the gap. Why the index decides suitability: too broad and the hedge carries basis risk against an average the operator never touches, too narrow and it is too thin to settle against. Capturing private cloud clears without losing the specific cluster is the whole problem. That index is the part we built.

A month ago I wrote that two layers were missing under the procurement paradigm: a way to hedge forward compute price without buying more capacity, and a benchmark worth settling it against. Today the first one ships. Architect is listing H100 through B300 futures, settled to our index. The contract was the easy part. The index is the work: broad enough to catch where compute actually clears in private deals, narrow enough that the hedge tracks your cluster and not a market average. Now there is a price to trade around.

Architect @Architect_Fi

3 weeks ago

Compute futures indexes need to be broad enough to capture private cloud transactions but narrow enough to minimize basis risk for datacenter hedging. Architect’s Nvidia H100, H200, B200, and B300 futures will offer the best of both worlds on the American Innovation Exchange.

2 7 36 21K 8

Interesting piece from @streetwiseprof. The implicit dichotomy in the piece is index-first vs spot-first. What's actually happening is both layers being built at once. Standard deliverables, settlement conventions, transaction-data clearing are under construction in parallel with the index launches. The gating step isn't sequencing. It's whether the plumbing converges on the same conventions across launches.

Useful breakdown. The round-trip part lands harder once you read the cash flows as a prepaid forward. Buyer prepays the producer to lock physical output, producer keeps residual ownership, end-user gets the capacity. Oil and gas has done this for forty years and the IFRS / VIE machinery already handles it. The Valor structure rhymes with that template. The accounting issue is which standard the audit chooses to apply.

Bond issuance is one channel. The other shows up in private credit. CoreWeave DDTL 5.0 sits at Ba2 / BB+, well outside investment grade. Cerebras took $1B from OpenAI as a loan that repays in compute, not cash, which is a prepaid forward in everything but name. Total compute being financed is meaningfully larger than the bond tape suggests.

Ensembling four listed indices into one composite gets you a number, not a price. None of the four indices above are based on transacted data at the volume layer where buyers actually clear. Averaging four reference quotes does not solve the underlying problem: there is no settled benchmark for a broker to take spread against. The broker layer you describe needs a clearing reference first.

Four published H100 rental indices. Same chip, eight months. The bear-market read this week leans on one index pulling back from a spike. The others don't show a reversal. There is no GPU price. Calling a single index move a market move makes the new market harder to form.

The option Zuck describes only has value if there is a market on the other side. Right now there isn't, at the relevant scale: the xAI-Anthropic deal in the S-1 cleared bilaterally on direct phone lines because there was no venue. If Meta ever pulls the trigger, the deal type they'd run is the same residual-value-swap shape, not a spot rental.

The S-1 disclosure makes the residual-value-swap shape explicit. xAI's 11% MFU on Colossus 1 wasn't earning the depreciation curve, so they passed the in-period utilization risk to Anthropic at full capacity and kept the residual. Appreciating in the spot tape because the rental short end can't clear; the depreciation curve still runs under the hood.

Labs with under-utilized compute is a narrow population. Three or four operators at $B scale can sign bilateral phone-call deals with each other. Below them, the under-utilization problem is real and the matching plumbing is missing. PPA era at the top of the stack, where two CFOs can dial each other; nothing yet below it, where the bilateral phone-call market does not reach. That gap is the structural opening.

Bulls who call on demand under 5 percent are reading a snapshot, not a curve. The whole point of a rental short end is that it carries the hedging load when no forward venue exists. NBIS at 30 percent mid cycle is what a forward curve looks like when it has to live inside the spot market.

The wrapper is filing ahead of the underlying. $GPUX and the ProShares filing both reference compute futures contracts, but there is no settled benchmark for B200 or B300 today, only listed-rate distortion and bilateral private prints. The contracts will need a reference layer before they can be productized at scale.

Token futures and compute futures are not symmetric. GPU-hours are a defined physical unit even if the benchmark on top is unbuilt. Tokens are not fungible across models, or even across input / output / cached within one model. Token futures have a unit problem one layer deeper than compute futures do. The wrapper layer is racing both. Wrote up the compute side of the gap: computedesk.substack.com/p/cerebras-is-…

Thank you Lazarus. The $GOOG side of the trade has a quieter read. Carving an external venue is the only way to discover how much TPU capacity sits outside their preferred internal allocation. Self-use was opaque. A 5B-backed venue runs the price test in public. The $BX side is just the financing partner.