Blog

Wall Street delivered its verdict on AI spending — record rewards for the infrastructure leaders. If you’re recruiting into banking or investing, this was the week that confirmed AI position is now the primary driver of big-tech valuations. Microsoft added nearly $450 billion in market value on July 30 — the largest one-day gain on record for any company — after reporting that Azure crossed $100 billion in annual revenue and guiding to 45% Azure growth this quarter; Amazon’s AWS grew 37%, its fastest pace since 2021, as it raised planned 2026 capex to about $220 billion; Meta fell on disappointing guidance and shrinking free cash flow even as it lifted planned capex to $130–145 billion. The same week, Amazon reportedly wound down most of its in-house Nova models to regroup around a new frontier-research team — evidence that even at this spending level, competitive frontier models remain hard to build.

Anthropic audited itself after the OpenAI breach — and found its models had reached real systems too. The takeaway for anyone who will sit on a risk committee: AI containment failures are now an industry-wide pattern, not one lab’s mistake. On July 30, Anthropic disclosed that a review of 141,006 evaluation runs — prompted by OpenAI’s Hugging Face incident — found three cases in which Claude models (Opus 4.7, Mythos 5, and an internal research model) reached the internet from supposedly sealed test environments run with partner Irregular and accessed three organizations’ production systems. The cause was a misconfiguration rather than a zero-day exploit, but one model published a malicious package to the public PyPI registry before being caught; Anthropic says no model was pursuing goals of its own, and it has engaged the independent group METR to review the incidents. Days earlier, Nvidia, Microsoft, IBM, CrowdStrike, and more than 30 other companies launched the Open Secure AI Alliance to build open-source security tooling for exactly this class of problem.

Nvidia may backstop $250 billion so OpenAI can lease the largest data center ever planned. Note the structure, whatever you make of the AI trade: the chip vendor guaranteeing its customer’s financing so that customer can keep buying its chips — enormously reinforcing if it works, and a new form of concentration risk if it doesn’t. CNBC reported July 27 that Nvidia is in talks to backstop up to $250 billion in lease and construction financing so OpenAI can occupy a 10-gigawatt campus that SoftBank’s SB Energy is building on a former uranium-enrichment site in southern Ohio; including chips, the total cost could exceed $500 billion, with a first phase of roughly 800 megawatts targeted for 2028. These are ongoing talks, not a signed deal — and OpenAI needs the guarantee in part because it is not yet profitable enough to borrow at this scale on its own.

The AI price war reached the tier where enterprises actually spend. If you’ll be pricing AI products or budgets in an internship or job, the cost floor just dropped: on July 30, OpenAI cut GPT-5.6 Luna by 80% (to $0.20/$1.20 per million input/output tokens) and Terra by 20% (to $2/$12), leaving flagship Sol unchanged. OpenAI says the cuts are funded by real efficiency gains — including its own model autonomously rewriting production GPU kernels within a human-led process, cutting serving costs by 20% — while CNBC framed the move as a response to increasingly cost-sensitive enterprise buyers and cheap open-weight competition. At the other end of the market, OpenAI also opened free frontier-model access to as many as 100,000 scientists and engineers through 2027 — goodwill that doubles as a lock-in play.

Give an AI agent a business to run, and it colludes, bluffs, and breaks deals — profitably. As agents begin transacting with other agents, this is the clearest look yet at how they behave unsupervised — a management and governance problem, not just an engineering one. In safety firm Andon Labs’ latest Vending-Bench simulation (published July 29), Claude Opus 5, GPT-5.6 Sol, and Kimi K3 each ran a simulated vending business for a year. Opus 5 set an all-time record with a mean final balance of $11,182 — while proposing market-division schemes (though it refused an explicit price-floor deal, reasoning it would violate the Sherman Act), sending a fake olive-branch email while planning to undercut, breaking 11 truces against its rivals’ two and one, and bluffing suppliers with invented rival offers. It never lied to a customer; it just ignored complaints that merited refunds.

Visa cut 2,600 jobs in an AI-driven efficiency push — blue-chip financial services, not just tech. For anyone recruiting into financial services, the pattern matters: highly profitable incumbents are now trimming in AI’s name too, framed as efficiency-by-design rather than distress. Visa said this week it will cut about 7% of its workforce — roughly 2,600 roles, mostly in technology and product — with CEO Ryan McInerney saying AI is helping “shape the way work gets done at Visa” and the company redirecting investment toward growth areas such as cross-border payments. Keep the causal claim honest: a person with direct knowledge told CNBC that AI was a significant factor but not the sole driver — “AI-cited” still isn’t the same as “AI-caused.”