Founder and CTO: @ComputeDesk. Exited Pace Revenue to FLYR. Formerly VP of Research at Winton, particle physics researcher at Harvard and CERN
Joined July 2018
- Tweets 165
- Following 74
- Followers 1,058
- Likes 66
16 Photos and videos
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/e…
1
1
9
336
David López Mateos retweeted
Jun 4
"The suitability of the Hedge depends on the index."
@Architect_Fi compute futures powered by @ComputeDesk indexes.
Jun 4
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 ↓
7
14
679
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.
Jun 4
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
25
1,655
May 28
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.
1
2
7
2,668
May 19
Cerebras 2025 GAAP profit was $238M. Strip out the $363M non-cash gain from extinguishing a forward contract liability and the operating business loses $76M.
The auditors classify these prepaid capacity commitments as forward contracts because that is what they are.
New piece on commodity finance applied to compute: computedesk.substack.com/p/c…
1
5
492
May 14
Update on the Cerebras open. Book at 185 implied 56B. Tape opens at 365, that's about 100B. The institutional book could not price the OpenAI 750MW capacity strip or the 1B working-capital prepaid forward. The public tape did, in the first hour. This is what price discovery on bilateral capacity commitments looks like when it finally goes multi-venue.
1
6
424
May 14
Cerebras prices at 56B. The most under read part of the S-1 is the OpenAI deal. 750MW of inference capacity is a strip of long dated forwards. The 1B working capital loan that repays in compute delivery is a prepaid forward on capacity, the same template oil and gas producers used to fund themselves through VPPs. This is commodity finance applied to compute.
8
375
May 13
CME is launching compute futures this year. That is the layer everyone has been arguing for. It only works if three pieces exist: standardized contract specs, transparent transaction data, a benchmark the contracts settle against. Without those, the futures price drifts from where compute actually trades. The data layer is the prerequisite, not the product.
1
7
391
May 12
1/ Two GPUs from the same generation. The price relationship between them has flipped four times in two months.
Listed prices say B200 and B300 cost the same. The actual trades disagree, and they keep changing their minds.
1
12
648
May 12
3/ B300 only started trading in late February. Two months in, its price range is as wide as B200's was in its first months. That's not unexpected for a new chip.
What's more interesting: even mature B200 still clears with a wide range. The market hasn't converged.
1
2
185
May 12
4/ Listed prices say the market is settled. Transaction data says it isn't.
Supply is growing. Prices aren't converging.
Until B300 finds its level, watch where it actually trades. We publish those rates.
Full piece: computedesk.substack.com/p/w…
1
100
Anthropic-Colossus isn't a compute deal.
xAI ran Colossus 1 at ~11% MFU. A 5-year offtake transferred utilization risk to Anthropic.
That's a residual value swap. Except no market for one exists in GPUs, so they hand-built it bilaterally.
4
339
This is xAI rotating training to Colossus 2 and finding Anthropic has more inference demand than Grok does. Otherwise xAI would keep it for themselves. Compute allocation is a market-share signal. The buyer with the bigger product wins the prior-gen rig.
The most interesting dynamic in AI right now:
Anthropic just rented the entire compute capacity of SpaceX / xAI Colossus 1.
220,000 NVIDIA GPUs. 300 megawatts. Online within the month.
Cursor also has a SpaceX / xAI compute deal... and SpaceX reportedly has a $60B option to acquire it.
What changes for Anthropic today:
> Claude Code limits doubled
> Peak-hour restrictions removed
> Opus API limits raised
Some of the companies competing with Elon are now running on Elon's supercomputer.
Wild.
5
519
1/ Securing GPUs today means owning them. Multi-year reservations, 20% prepayment, one chip family, one provider. Buyers do this because there is no alternative.
3
2
13
1,736
3/ Two things are missing, not one. A liquid market for short-tenor GPU compute. A benchmark with hedging instruments on top. They reinforce each other but don't strictly require each other (h/t @0xmoonrunner whose recent participant map sets this up).
1
1
240
4/ Wrote up the full argument, including why hardware generational refresh strengthens the case rather than weakens it: computedesk.substack.com/p/t…
148
How VRAM actually gets consumed serving DeepSeek V4-Pro at FP4 across 4 GPUs.
4× B200, per card: weights overflow. Doesn't run. 4× B300, per card: 75 GB free per card for KV cache and concurrency.
Same hourly listed rate. Different operational reality.
1
5
550
Apr 30
What fits where, FP8 vs FP4 across B200 and B300.
At FP8 they look interchangeable. At FP4, B300 crosses thresholds B200 can't.
This is the chart that explains why B200 and B300 rent for the same listed price but aren't the same GPU. From Tuesday's analysis. Link in reply.
2
7
296