A large systems integrator just went all-in on Claude with production numbers to show for it, the hyperscalers keep raising their infrastructure budgets even after chip stocks wobbled, and the price of a token has collapsed to a point where the labs are openly competing on how cheap they can go. For a European buyer, the useful read is that capability is converging and getting cheaper — so the decision that's left is about total cost and where the model runs.
Cognizant goes “Global Premier” on Claude — with production ROI, not just a press release
Cognizant on 27 July became a Global Premier Partner in Anthropic's Claude Partner Network, deepening a relationship that began in late 2025. What makes this more than another consultancy-signs-a-lab headline is the operational detail: more than 30,000 associates have completed Claude training, with a target of 5,000 Frontier Certified Engineers and 10,000 Frontier Business Operators and roughly 40,000 professionals already in the certification pipeline. Cognizant is embedding Claude directly into its own industry platforms rather than treating it as a side experiment.
The reported production results are the part worth noting for anyone weighing an agent programme: contract review up to 40% faster with over 88% extraction accuracy, and underwriting research cut from hours to about a minute, in live deployments across manufacturing, life sciences, and insurance. That lands against a sobering counterpoint — Gartner projects over 40% of agentic AI projects will be cancelled by the end of 2027 on unclear ROI and weak controls. The two facts aren't in tension so much as a map: the value is real where the workflow is narrow, the data is structured, and someone owns the governance — and it evaporates where those conditions are missing.
The capex bill keeps rising — and Microsoft reports today
The infrastructure spend behind all of this shows no sign of slowing. Alphabet raised its 2026 capital-expenditure guidance to between $195B and $205B on 22 July — up from $180–190B — citing capacity shortages driven by stronger-than-expected cloud demand, with Q2 capex of $44.9B and a cloud backlog of $514B. Microsoft reports its fiscal Q4 results today, with roughly $190B in FY2026 capex and Azure growth guided near 39–40% in constant currency — the numbers investors are watching most closely.
The context is that this spending is still accelerating even after last week's wobble, when the Philadelphia Semiconductor Index slid into a technical bear market on doubts about whether the build-out will earn its return. For European teams the ticker matters less than the geography: the compute being committed is overwhelmingly US-hosted, which keeps jurisdiction and sovereignty squarely on the procurement checklist regardless of how the earnings prints land.
The token-price war is now the main event — and no one's obviously winning it
While budgets climb, the price of inference is falling off a cliff. By some estimates the cost of equivalent intelligence is down 90% or more over the past year, and the labs have stopped being coy about it: on 14 July Sam Altman publicly challenged Anthropic, saying OpenAI would serve GPT-5.6 Sol at a quarter of Claude's flagship price. The economy tiers tell the story: OpenAI's Luna at roughly $1/$6 per million tokens, Google's Gemini 3.5 Flash near $1.50/$9, and Claude Sonnet 5 at an introductory $2/$10.
The uncomfortable subtext, as Sherwood put it, is that this is a price war “where no one is making money” — capex rising, per-token revenue collapsing, and margins caught in between. For buyers that's an opportunity with a catch: the marginal cost of capability keeps dropping, but the economics pushing it are unstable, and a model priced at an “introductory” rate today can reprice tomorrow. The durable variables in a procurement decision are therefore the ones the price war doesn't touch — where the data is processed, under whose jurisdiction, and how portable your stack is if the terms change.
Quick Hits
- EU AI Act GPAI enforcement goes live 2 August — In three days the Commission's enforcement powers over general-purpose AI models take effect, ending the one-year grace period. Under Article 101 the AI Office can request documentation, evaluate models, order mitigations, restrict market access, and fine up to €15M or 3% of global turnover.
- Alibaba's Qwen 3.8 Max stays a preview — The 2.4-trillion-parameter multimodal model shown at WAIC on 19 July remains a preview endpoint with no open weights, no license, and no release timeline — part of Qwen's broader tilt toward closed flagships.
- A genuinely open counterpoint: poolside Laguna S 2.1 — poolside shipped Laguna S 2.1 on 21 July with weights on Hugging Face day one under the OpenMDW-1.1 license — the inverse of the “announced but not downloadable” pattern that has defined much of the month.
- Gemini 3.5 Pro slips again — Google's release has reportedly slipped once more on coding shortfalls with no new timeline, even as pre-training on Gemini 4 has reportedly begun.
- Mistral's open-weight MoE still teased, not shipped — Arthur Mensch's “fat but sparse” model remains in partner early access with no parameter count, benchmarks, or license disclosed; a broader release is expected later this summer.