Earnings week turned into a referendum on the AI trade, and the verdict was specific: the market now pays for AI spending that shows up as cloud revenue, and marks down the rest. Amazon delivered the clearest proof point, while Brussels switched on the first user-facing rules of the AI Act, OpenAI made an extraordinary and unverified research claim, DeepSeek kept cutting the price of capable AI, and Google DeepMind reshaped the team behind one of science's most celebrated AI breakthroughs.
Amazon's $220 billion bet finally shows up in the cloud numbers
For a year the question hanging over the hyperscalers was whether record AI capital spending would ever translate into growth. Amazon's second-quarter results, reported July 30, answered it for now. AWS grew 37% to roughly $42.2 billion, its fastest expansion in eighteen quarters, with an AI and custom-silicon business running past a $25 billion annual rate. The stock rose around 9-10%, and the company raised its 2026 capital-expenditure guidance to about $220 billion, up from roughly $200 billion.
The caveat came from CEO Andy Jassy, who named soaring memory-chip costs as a genuine constraint even as demand outstrips supply. That single line connects the whole week: the bottleneck in AI is migrating from GPUs toward memory and power, and the companies that manage that supply chain best will set the pace.
The EU's AI Act starts talking to users
On August 2 the European Commission began enforcing Article 50, the transparency layer of the AI Act. Interactive systems must now tell people they are dealing with AI, and AI-generated or manipulated content, deepfakes included, must carry machine-readable marks and clear labels. The obligations bind both the providers who build systems and the deployers who put them in front of users, and there is no open-source exemption for these particular duties.
The practical effect is that content provenance and disclosure stop being optional polish. Any product serving EU users now needs watermarking and labelling built in, which quietly expands the market for the tooling that supplies it and turns a policy document into an engineering requirement.
OpenAI's big claim, and the case for waiting
OpenAI used a Washington audience to preview a next-generation model, tentatively called Astra, that coordinates sub-agents across long tasks. Its most striking assertion, reported by The Decoder, is that the system produced solutions to ten problems in mathematics and theoretical computer science that had been open for more than a decade.
If that holds, it would move AI from assisting research to closing it. The word "if" is doing real work. This is OpenAI's own account; independent skeptics dispute whether the proofs verify, the results are not settled, and even the name is provisional, with the model possibly shipping as GPT-6. The honest read is that long-horizon multi-agent reasoning is the capability everyone is racing toward, and OpenAI just staked a very public claim on it that the field now has to check.
DeepSeek keeps cutting the floor
While frontier labs chase reasoning, DeepSeek keeps attacking price. It re-post-trained its budget model, the 284-billion-parameter V4-Flash-0731, and reports that it now beats the company's own Pro flagship across all nine of its published agent benchmarks, including a Terminal-Bench 2.1 score around 82.7, at roughly $0.14 per million input tokens.
Those numbers are vendor-reported and sensitive to how the benchmarks are run, so independent evaluation still matters. The direction, however, is unambiguous: a cheaper model outscoring the maker's own premium tier keeps pushing the cost of capable agent work down, which widens what is economical to automate and pressures every provider selling a pricier tier.
DeepMind reshapes a Nobel-winning team
The people behind AlphaFold won the 2024 Nobel Prize in Chemistry. According to a Financial Times report, DeepMind has now folded that team into its Gemini organisation, and roughly a quarter of the original authors have left, several of them reportedly for Anthropic. Google characterises the change as a restructuring rather than a shutdown.
The verb matters less than the movement of people. When researchers who did foundational, prize-winning work leave for a frontier competitor, the AI talent market is behaving as the real constraint on scientific progress. Where that talent lands is a leading indicator of who builds the next wave of science-grade AI, and this week it pointed at least partly toward Anthropic.
The through-line is a market that has grown discerning. It is paying for AI capex that converts into revenue, as Amazon showed, while a structural memory crunch — DRAM prices up sharply and HBM effectively sold out — puts a ceiling on how fast anyone can scale. Watch three things next: whether OpenAI's Astra math claims survive independent verification, how regulators enforce the EU's new transparency rules in practice, and whether the memory bottleneck eases or forces another round of capex and pricing moves in the second half. Signal, not advice.