Price War in the Chip Aisle

Tue, 13 Oct 2026 09:00:00 -0700

Five frontier models now sit within five points of each other in terms of intelligence, but their cost per unit differs by 3.8x.

Google’s Argon prices at the bottom of the cluster at $1.99 a task¹, 74% below the most expensive model, Fable 5.1. It is also 67% below Opus 5.5, 56% below Muse Wonder & 39% below Astra.

If we are to believe Michael Mauboussin’s principle that prices convey information², then we can deduce that different models, architectures & different costs of GPUs translate directly here.

Why could this be?

Google has been building machine learning models in parallel to its chips for well more than 15 years, more experience than anyone else on this list³.

The evidence is in Google’s own peer-reviewed work on its chips, their Tensor Processing Units (TPUs). The fourth iteration TPU chips, released in 2023, ran 1.2x to 1.7x faster than the Nvidia A100 while drawing 1.3x to 1.9x less power⁴. At the low end of the throughput range against the midpoint of the power range, that yields 80% more tokens for the same serving cost⁵.

If that advantage has carried through to today, then Google could very well be serving tokens at 44% less than its peers, while operating at similar margins. It would also explain the disproportionate scale of Google’s data center buildout relative to its smaller cloud business⁶.

Google is no longer alone in this. Every lab that can afford custom silicon is now building it. Microsoft’s Satya Nadella, introducing Maia 300 in October 2026 :

“Starting next calendar year, we will scale our Maia family to multi-gigawatt deployments, making it the primary inference system for our own product and services. And we’re introducing Maia 300, which delivers 40% more tokens per watt compared to leading GPU-based systems.”⁷

OpenAI’s first inference chip, Jalapeño, went further. Tested on SemiAnalysis’s InferenceX suite against Nvidia’s GB200 & GB300, it delivered 1.5x to 1.9x more throughput per watt & 1.7x to 3.6x lower end-to-end latency :

“Jalapeño delivered a better combination of throughput, power efficiency, and latency. Although performance is sometimes reported per chip, we believe the more useful standard is performance per unit of power.”⁸

As intelligence per watt doubles every six months, pricing should continue to deflate in the world of tokens, with margins remaining in the 40-70% range, particularly for frontier models⁹.

Cost per task is the weighted average cost to run one Intelligence Index task, aggregating input, cache hit, cache write, reasoning & answer token prices across the ten evaluations in Artificial Analysis Intelligence Index v4.3.2, weighted by each evaluation’s index weight. Source : Artificial Analysis, retrieved October 2026. All models measured at maximum reasoning effort. ↩︎
Michael Mauboussin’s principle that prices convey information: the price embeds the market’s expectations for growth, returns & risk. The reversal method begins from the price, derives the expectations embedded in it, then compares those to your own. Source : Michael Mauboussin, Expectations Investing: Reading Stock Prices for Better Returns, expectationsinvesting.com. ↩︎
Google began its TPU project in 2013, the same year it projected that deep neural networks could double data-center compute demand, & deployed the first chip to its own data centers in 2015, more than a year before announcing it publicly in May 2016. The program has shipped seven generations since, reaching TPU v7 (Ironwood) in 2025. The 15-year claim counts from the project’s start against the labs this comparison covers, none of which had a comparable custom-silicon effort at that date. Source : Norman P. Jouppi et al., “In-Datacenter Performance Analysis of a Tensor Processing Unit”, ISCA 2017, & cloud.google.com. ↩︎
Norman P. Jouppi et al., “TPU v4: An Optically Reconfigurable Supercomputer for Machine Learning with Hardware Support for Embeddings”, ISCA 2023, arXiv:2304.01433. For similarly sized systems the authors measure TPU v4 at 1.2x–1.7x faster than the NVIDIA A100 while using 1.3x–1.9x less power. On comparison against the H100, the authors write that the 700W part “was not available at AWS, Azure, or Google Cloud in 2022” & that “the appropriate H100 match would be a successor to TPU v4 deployed in a similar time frame & technology.” Figures are Google’s own measurements, not independent verification. ↩︎
Tokens per GPU-hour dollar is the product of the throughput multiplier & the power multiplier, since power sets the hourly cost. The 80% figure uses the low end of the published throughput range (1.2x) against the midpoint of the published power range (1.5x), giving 1.2 × 1.5 = 1.80. Applied across the full published ranges the same calculation spans 1.56x to 3.23x, or 56% to 223% more tokens per dollar. The estimate therefore rests on the midpoint of a range, not a measured serving result, & assumes the v4-era advantage persists across subsequent generations & that both chips carry comparable total cost of ownership at the system level. ↩︎
Google is outspending Microsoft on capital expenditure despite running a cloud business roughly 37% its size. Google raised full-year 2026 capex guidance to $180–190b against Microsoft’s roughly $120b, & Alphabet financed part of the gap with a $32b debt offering, including a 100-year bond. Source : The $112 Billion Quarter, April 2026, drawing on Alphabet, Amazon & Microsoft Q1 2026 reporting. ↩︎
Satya Nadella, post on X, October 12, 2026. Maia 300 is Microsoft’s third-generation custom AI accelerator, scheduled for data centers in 2027, at up to 96 chips per rack. Microsoft has not published an independent benchmark, & the 40% figure is company-reported against an unnamed set of “leading GPU-based systems,” presumably Nvidia. Nadella also said Maia 200 delivers up to 1.7x more tokens per watt in Teams & Copilot, without a stated comparison basis. ↩︎
OpenAI, “Jalapeño’s first results show industry-leading speed and efficiency in AI inference”, August 25, 2026, reporting results measured on SemiAnalysis’s InferenceX suite with OpenAI engineers present, across GPT-OSS 120B, DeepSeek R1 670B & Kimi K2.5 1T. OpenAI normalized to published package power & reports Jalapeño’s sustained draw at or below 550W against a 700W rating. Three caveats apply : the comparison is against Blackwell-generation GB200/GB300 rather than the newer Vera Rubin platform, which also uses HBM4 & began shipping to customers in 2026, making Rubin the more like-for-like test ; Jalapeño ran single-token prediction while some Nvidia baselines used multi-token prediction, which narrows the advantage to roughly 1.5x on a matched basis ; & the benchmark measures performance per unit of power, not total cost of ownership, excluding host processors, cooling, networking, software & utilization. SemiAnalysis participated in the testing & called the Blackwell comparison “somewhat incomplete and unfair” for that reason. ↩︎
Amin Vahdat, Google’s chief technologist for AI infrastructure, on the Sequoia Capital Training Data podcast with Sonya Huang, October 2026 : sequoiacap.com. Vahdat says most of the gain in intelligence per watt comes from model-side improvements rather than silicon, & that his team has been asked to roughly double effective serving capacity, measured in token generation rather than raw FLOPS, every six months. The episode chapter is titled “Doubling token capacity every six months, & where the gains actually come from.” The 40-70% margin range is this essay’s estimate, not his, & the direction of token pricing is an expectation rather than a measured forecast. ↩︎

The Theory 2026 AI Survey Results

Tue, 13 Oct 2026 08:00:00 -0700

The 2026 Theory AI Survey surprised us, just as the 2025 edition did. In 2025, we found that context, limited AI performance. In 2026, that seems to be solved.

455 senior technical builders replied to give us a cross-section of the AI ecosystem today.¹

Security is the biggest barrier to adoption

Security rose from sixth to first among pain points in twelve months. 43% of respondents rate security & governance the most important pain point, & evals come in second at less than half, 22%.

Only 10% say security tooling delivers the most value today. 18% expect it to lead in the future, the largest expected gain of any area.

Nearly half of teams route queries

Cost optimisation as well as model selection have become standard.

45% of teams now route queries across models, 23.7% on a router they trained & 21.5% on a proprietary one, while 54.7% send every query to a single model.

Cloudflare dominates sandboxes

Of the 259 teams writing code with agents, 62% host their sandboxes on Cloudflare, three times the closest competitor. Then Other at 17%, Vercel at 17%, Modal at 16%, E2B at 12%, Daytona at 9% & Sail at 7%. The one control a builder cannot fake is concentrated in a single utility, & it is billed as compute rather than as security.

BI performs more than half of agent analysis

56% of respondents use business intelligence tools to analyse the performance of their agents. The surprising statistic is not far behind : coding agents also perform some of the work, at 48%. Both systems are excellent at parsing agent behaviour. A roughly equivalent number use dedicated observability tools, at 41%, which does suggest a convergence of the category, at least for this one use case. Spreadsheets, that tardigrade² of enterprise software, continue to survive at 30%.

There is a lot more to unpack within the full survey, including open versus closed source model adoption, the rise of MCP despite doubts last year, & the still early adoption of evals. Overall, it is clear how quickly the ecosystem is evolving, & how much sophistication has permeated the enterprise, more quickly than any wave we have seen.