Why Compute Prices Are Still Rising | Steve Hou of Silicon Data on GPU Rental Indices, Fed Meeting, and Anthropic S-1 (LIVE)

Silicon Data’s usage-weighted index of prices across open and closed AI models has fallen since the end of May even as token usage rose substantially. Steve Hou said that can indicate commoditization, but it can also reflect AI spreading to more users who choose cheaper models and generate more tokens, as well as substitution from expensive proprietary models to open models. By contrast, rental rates for H100, B200 and B300 GPUs have surged in recent weeks, especially for the most advanced chips. Rising on-demand rates across tracked GPUs point to tight supply, but do not show whether demand comes from a broad base of enterprises or includes circular spending by AI-native and AI-adjacent companies.

The price chart uses separate scales: token prices are dollar-normalized per million tokens, while GPU rental prices are indexed to 100 on May 31. Hou stressed that it does not show token-production costs exceeding selling prices; newer chips serve tokens efficiently, and serving-token margins remain positive. He contrasted the current mix of rising rental prices and falling token prices with Q2, when token prices and GPU rental rates were both high, but rentals had not yet surged as much. Silicon Data plans to develop compute futures that could let suppliers hedge against falling prices and revenue, but the contracts still require CFTC approval. Its current rental term curves, which extend from on-demand to 36 months, are not traded futures; longer-term data is thin, and Hou did not definitively classify the curve as contango.

Anthropic’s S-1 showed extraordinary enterprise adoption and strong Q2 results, including Opus’s popularity. Although the filing disclosed slower Q3 revenue growth, Hou said Ramp evidence pointed to a possible reacceleration tied to Opus, while emphasizing that an inflection was not yet established. Q2 results implied an 88% gross margin, or about 82–83% after cloud-provider shares, excluding training costs; the reported operating margin was 8%. The discussion cautioned that reported losses of $72 billion or $80 billion were misleading: an Amazon convertible note was treated as a derivative liability whose value rose with Anthropic’s market capitalization. Hou called Q2 growth and profitability strong, but said future expectations depend on whether growth and competitive advantage persist; Anthropic’s filing itself says its current growth rate will not be sustained.

Enterprise adoption and retention may leave Anthropic room to grow: Gavin Baker was cited as saying half of the Fortune 500 had adopted it, while Anthropic accounted for 0.1% of their expenses. Onboarding, compliance and a seamless user experience could make the product harder to replace, though speakers disagreed on whether agents will make customers more model-agnostic and weaken that advantage. Chinese open models were described as charging 5–10% of Anthropic’s prices, but speakers questioned whether benchmark results translate into comparable real-world performance, particularly for coding agents. Lower prices could matter to enterprises at scale, but the extent of switching remains uncertain. The bear case includes token-price deflation, Chinese competition and the possibility that the trillions needed to remain competitive could bankrupt companies or burden the US government.

Hou said data-center investment ultimately depends on evidence that AI is producing returns through productivity, profitability, revenue growth or broader economic growth. Power and land access, uncertainty over future capacity needs, and higher memory, power and construction costs also shape buildout and capex. He viewed widening AI-related credit spreads as signs of doubts about investment returns, but said interest rates alone would not necessarily stop investment. At the Fed discussion, Chairman Warsh said officials were trying to tighten financial conditions and welcomed markets doing so; Hou said the greater risk would be AI-driven inflation prompting the Fed to slow the economy and alter buildout plans. Falling valuation multiples for memory, storage and Nvidia likewise suggested investors doubted the durability of current earnings and rental prices.

Google Search grew 12%, Google Cloud grew more than 100%, and Alphabet’s operating margin rose from 23% a year earlier to 36%. Speakers described Cloud growth as profitable and including TPU sales; they attributed Alphabet’s miss to non-AI business, particularly Search, rather than its fast-growing AI-related Cloud operation. The host noted that Google and Microsoft reported cloud growth while Q3 capex came in below expectations, but neither had cut capex and Microsoft’s call had not yet begun. Oracle forecast 35% revenue growth through 2030 and hinted at greater cash-flow discipline, which could weigh on the AI hardware trade. Hou said cloud providers that sold capacity forward gained more predictable revenue, subject to tenant credibility, but gave up upside as spot prices rose. CoreWeave had positive margins on existing data-center activity, while new investment could leave cash flow negative and require borrowing; the host said it had locked in compute-sale agreements for three, five and even seven-plus years before the price surge.