InfrastructureCompute, chips & cloud
SF Compute CEO Argues GPU Futures Need Physical Delivery as CME's Cash-Settled Contracts Await Regulatory Review
SF Compute CEO Evan Conrad argues GPU futures should be built on actually delivered capacity, because compute varies too much by hardware, storage, networking and interruption terms for thin cash-settled indexes to track prices buyers can really get.
From the Latent Space podcast episode of October 20, 2026 and cited sources: Conrad does not oppose cash-settled compute futures. [1] [3] [7]
SOURCES[1] The Programmable Data Center — Evan Conrad, SF Compute… latent.space[3] CFTC Prepares Review of AI Compute Futures letsdatascience.com[7] Compute futures: CME GPU contracts and AI buyers (Spheron) spheron.network
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- Latent Space podcast episode, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 20, 2026 podcast episode published
- Covered
- Oct 21 edition
- Evidence
- 8 sources across 8 sites; transcript sourced
- 1The Programmable Data Center — Evan Conrad, SF Compute…latent.space
- 2San Francisco Compute sells $245M worth of computesfcompute.com
- 3CFTC Prepares Review of AI Compute Futuresletsdatascience.com
- 4Jump Trading and DRW commit traders and capital to compute…cryptobriefing.com
- 5Jump Trading Plans Market Making for CME and ICE Compute…tokenpost.com
- 6SF Compute — AI Engineerai.engineer
RisingHow it’s measured
YouTube video · 29.3K views · 118 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
Conrad does not oppose cash-settled compute futures. He argues they should be based on prices from physically delivered capacity. Otherwise, indexes built on small amounts of data could be manipulated as LIBOR was, and hedges could drift away from what capacity actually costs. He sees cash-settlement firms such as Silicon Data and Compute Exchange as likely partners, not rivals. [1] [3] [7]
SF Compute has not set standard contract grades. Its matching engine pairs orders that each carry several conditions, such as time window, chip generation, storage that affects recovery after an interruption, and NVLink domain size. Conrad says two VM types from the same provider can differ by about 40% in GPU kernel-launch time. The cost of this design is wider gaps between buy and sell prices. [1]
The market is still thin. About 5% of SF Compute's capacity is on its order book, and buyers get back roughly $25 of every $100 spent by reselling time they don't need. Both are company-reported figures. The company also announced $245M in 36-month B300 cluster contracts with an unnamed AI company. [1] [2]
Conrad estimates that a B300 rents for about $5–6 an hour but can earn about $40 an hour selling inference at full use, or about $25 an hour at half use. He argues that one of those prices has to move. [1]
CME's cash-settled GPU contracts await regulatory review, and trading firms such as Jump and DRW are committing capital.
Read the full assessment
If Conrad's points about hardware differences hold, hedges tied to rental indexes may not track capacity buyers can actually get. The CFTC has raised similar manipulation concerns.
Executive brief
SF Compute CEO Evan Conrad argues that a compute futures market will be easy to manipulate unless it rests on GPUs that are actually delivered (physical settlement). He made the case on the Latent Space episode of 20 October 2026, while CME's GPU futures, which settle in cash only, were still waiting on regulators. His most striking figure: a B300 rents for about $5–6 an hour but can earn about $40 an hour selling inference at full use. Conrad says one of those prices has to move. SF Compute also reports $245M of B300 contracts. Jump Trading and DRW are now putting traders and capital into this market.
What changed and event timeline
CFTC opens comment period
The regulator asked for public comment on derivatives tied to computing capacity. It named liquidity, manipulation, customer protection and perpetual futures as concerns, with comments due Oct 20 ().
SF Compute's $245M B300 deal
An unnamed "leading AI company" signed two 36-month agreements, each worth $122.5M, for dedicated B300 clusters ().
Jump builds a compute desk
Jump hired the founding team of Itô Markets to trade bilateral compute contracts. DRW is also active ().
CME delays, Jump plans market making
CME's review now runs to Nov 9. CME has started distributing its H100 and B200 index data, and Jump plans to make markets in both the CME and ICE/Ornn contracts ().
Conrad interview published
Conrad describes SF Compute as a "programmable data center" and explains why it insists on physical settlement ().
Capabilities and access
- Products: virtual machines, bare metal, and managed Slurm and Kubernetes.
- Order book: a buyer can price anything from the next hour out to three years (~15:54).
- Flexible orders: a buyer can set a rule such as "any Blackwell under $6/hr" and be matched to either GB300s or B300s.
Read the full section
- Products: virtual machines, bare metal, and managed Slurm and Kubernetes. Reservations run from one node for an hour to thousands of nodes for years (AI Engineer).
- Order book: a buyer can price anything from the next hour out to three years (~15:54).
- Flexible orders: a buyer can set a rule such as "any Blackwell under $6/hr" and be matched to either GB300s or B300s. Named custom SKUs are new (~22:31).
- Self-serve networking: customers can split up their own InfiniBand network without filing a ticket (~35:06).
Technical analysis for researchers and developers
- No fixed grades: SF Compute did not define standard contract grades. This trades deeper liquidity for wider bid-ask spreads (~08:42).
- Pricing dimensions: interruption notice (minutes for batch jobs, months for live inference), shared storage such as Weka or VAST, connections between clusters, NVLink domain size, VRAM, and when the hardware was deployed.
- Same GPU, different speed: two VM types from one provider can differ by about 40% in GPU kernel-launch time, which lowers training and inference efficiency (MFU) (~07:31).
Read the full section
- No fixed grades: SF Compute did not define standard contract grades. Its matching engine pairs orders that carry several constraints (time window, hardware, how storage handles preemption). This trades deeper liquidity for wider bid-ask spreads (~08:42).
- Pricing dimensions: interruption notice (minutes for batch jobs, months for live inference), shared storage such as Weka or VAST, connections between clusters, NVLink domain size, VRAM, and when the hardware was deployed. Early H100 and GB200 units had firmware faults (~09:53–13:30).
- Same GPU, different speed: two VM types from one provider can differ by about 40% in GPU kernel-launch time, which lowers training and inference efficiency (MFU) (~07:31).
- Unverified: no benchmarks or measurement method have been published for any of these figures.
Claims and evidence
- Liquidity: 5% of SF Compute's capacity is on the order book. Buyers get back about $25 for every $100 spent by reselling time (~05:45).
- Volume: "several hundred million dollars" closed (~34:30). This fits the $245M announcement, but the buyer is not named.
- Price shape: prices dip close to delivery, rise, then fall further out. Conrad calls it a "hump," and it is anecdotal (~18:16).
Read the full section
- Liquidity: 5% of SF Compute's capacity is on the order book. Buyers get back about $25 for every $100 spent by reselling time (~05:45). The company repeats the 25% figure in its own release. Both come from the company.
- Volume: "several hundred million dollars" closed (~34:30). This fits the $245M announcement, but the buyer is not named.
- Price shape: prices dip close to delivery, rise, then fall further out. Conrad calls it a "hump," and it is anecdotal (~18:16).
- Arbitrage: the $5–6 rental versus ~$40 inference comparison is Conrad's own estimate (~19:28).
Context and prior work
- Cash-settled benchmarks: CME's contracts reference Silicon Data's rental indexes. Each covers 730 GPU-hours, with monthly contracts listed 36 months out (Spheron).
- DRW's role: DRW incubated both Silicon Data and Compute Exchange (Crypto Briefing).
- Other exchanges: ICE (with Ornn) and Nodal are also developing compute contracts.
Read the full section
- Cash-settled benchmarks: CME's contracts reference Silicon Data's rental indexes. Each covers 730 GPU-hours, with monthly contracts listed 36 months out (Spheron).
- DRW's role: DRW incubated both Silicon Data and Compute Exchange (Crypto Briefing). Conrad calls them likely partners: they do cash settlement, SF Compute does physical (~27:23).
- Other exchanges: ICE (with Ornn) and Nodal are also developing compute contracts.
Limitations, safety and contested findings
- Manipulation risk: Conrad compares thin indexes to LIBOR, which was manipulated, and warns of basis risk, where a hedge stops tracking the price of capacity you can actually buy (~26:12).
- Counterpoint: cash-settled futures protect against price moves but do not guarantee access to hardware (Spheron).
- Conflict of interest: physical settlement is SF Compute's own business model.
Read the full section
- Manipulation risk: Conrad compares thin indexes to LIBOR, which was manipulated, and warns of basis risk, where a hedge stops tracking the price of capacity you can actually buy (~26:12). The CFTC shares this concern where supply is concentrated and data is limited (Let's Data Science).
- Counterpoint: cash-settled futures protect against price moves but do not guarantee access to hardware (Spheron).
- Conflict of interest: physical settlement is SF Compute's own business model.
- Neutrality: the company says it owns no hardware so it stays neutral (~29:10). No independent party has verified this.
Business and practitioner implications
- Owners: clusters are financed over three to five years.
- Buyers: workloads that can be interrupted, and can run on several hardware types, can earn the spreads. Live inference pays for longer interruption notice.
- Finance teams: watch CME's Nov 9 decision. Hedging with cash-settled contracts carries basis risk until physically delivered prices feed the indexes.
Read the full section
- Owners: clusters are financed over three to five years. SF Compute offers to run the supercomputing layer, carry the SLA, and resell unwanted capacity (~36:18).
- Buyers: workloads that can be interrupted, and can run on several hardware types, can earn the spreads. Live inference pays for longer interruption notice.
- Finance teams: watch CME's Nov 9 decision. Hedging with cash-settled contracts carries basis risk until physically delivered prices feed the indexes.
- Sovereign AI: SF Compute pitches itself to countries building national clouds. Conrad said partner announcements would come "in a week or two" (~32:45).