AI Pulse Daily Brief | 2026-07-31
Reading time ~4 mins
The EU opened procurement for up to seven large AI compute sites, with bids due 12 November.
OpenAI cut the price of its cheapest model tier by 80 percent.
Two new OECD studies land together: firm-level data says the AI advantage follows complementary capability rather than model access, and every country surveyed now expects a documented risk check before AI touches the workplace.
Ed Zitron presses the one number no large cloud provider discloses.
Perspectives
Ed Zitron argues the AI build-out still lacks a disclosed revenue figure to justify it. Skeptic
Ed Zitron published a long essay on 28 July on the economics of the AI build-out. He reports that the four largest cloud providers added more than 850 billion dollars of property, plant and equipment since March 2022, while none of them discloses a directly attributable AI revenue line. He reads the pattern of supplier-backed financing and customer concentration as evidence that announced capacity has outrun proven end-customer demand. The essay offers no audited deal-level data, and the same figures fit a normal build-out whose revenue arrives after construction. What survives the argument is the disclosure gap itself, which sits outside the capability and availability criteria that foundation-model and cloud renewals are scored on today.
Netherlands & Sovereignty
The EU opened procurement for up to seven large AI compute sites, closing 12 November. Authority
The EU's joint high-performance computing body opened its call on 30 July for consortia to build and run up to seven sites, branded AI Gigafactories, in member states, including facilities split across borders. Proposals close on 12 November 2026, selection is expected in early 2027, and the announcement targets operation within 18 months of that. The Union and participating states will jointly buy compute access time rather than only fund the hardware, a structure meant to unlock more than 20 billion euros of private investment. Allocation rules and data-residency terms go into consortium bids this autumn, which is when the question of whether regulated workloads can run on that capacity is effectively settled.
European High-Performance Computing Joint Undertaking
Innovation
OpenAI cut the price of its cheapest GPT-5.6 tier by 80 percent. Vendor
OpenAI updated its GPT-5.6 product page on 30 July, cutting the price developers pay to use its cheapest model tier by 80 percent and its mid tier by 20 percent. The same announcement says the family is generally available and describes a test capability for running several AI agents together. The cut lands on the lowest-capability tier rather than the models most bank workloads would use, and no commitment period is stated. What it changes is the binding constraint on high-volume internal agent pilots, since work refused last quarter on running cost now stands or falls on approved-model, retention and regional-processing review.
Research
OECD firm-level data finds the AI advantage tracks existing capability, not model access. Institute
A 75-page OECD working paper published on 30 July combined official company surveys in France and Portugal, Portuguese employer-employee records, global AI patent filings and start-up data. Firms using earlier, non-generative AI between 2011 and 2022 were larger and more productive, but did not charge substantially higher margins than non-users. The generative wave looks different: the paper's early evidence points to benefits accruing to firms that already hold the data, skills and processes to absorb them. Concentration in AI patenting also correlates with higher sales concentration, and AI use reached 20.2 percent of OECD enterprises in 2025. The finding cuts against the assumption that buying the same models as a competitor buys the same position.
OECD: Competition in the age of AI: Initial evidence from microdata
Every country OECD surveyed now expects a documented risk check before AI touches the workplace. Institute
A separate 64-page OECD review published on 24 July mapped workplace AI measures across the G7, the European Union and four Latin American economies. It reports that governments are mostly applying existing labour, privacy and safety law rather than writing new AI-specific rules, and that every country surveyed recommends or requires a risk assessment. Practical guidance on transparency, explanation and accountability, by contrast, is still mostly stated as principle. Reskilling programmes concentrate on technical skills, with few measures aimed at staff who keep their jobs but see the work change. The gap matters because the surveyed expectation has moved from whether an employee-facing tool is allowed to whether its operator can evidence oversight on request.