AI News Flash · Week in Review

TSMC's Q2 blowout raises full-year AI capex to $64B, then falls 5%

By and · Editorial team, AI News Flash

The five stories that defined the week

TSMC's Q2 blowout raises full-year AI capex to $64B, then falls 5%

TSMC reported Q2 2026 net profit surging 77% year-over-year to a record high, with revenue of $40.2B at the top of guidance and a 67.7% gross margin that beat the ceiling of its own forecast. The more consequential number was the capex revision: full-year spend was raised from $52–56B to $60–64B, and the company announced an additional $100B Arizona commitment, bringing total US planned outlay to $265B. CEO C.C. Wei raised the full-year revenue growth target to 'slightly above 40%,' a significant step up from the prior 30%-plus guide, and confirmed 3nm lines are running above 100% utilization with CoWoS packaging remaining the industry's primary bottleneck. The stock fell 5% anyway — a signal the market had already priced perfection and is now discounting AI infrastructure stocks on the gap between today's capex and when it converts to free cash flow. Watch whether the margin compression from the 2nm ramp (guided at 65–67% gross margin in Q3, down from 67.7%) turns into a sustained trend or a one-quarter absorption cost; the answer will determine whether the AI capex cycle thesis remains intact or starts cracking at the foundry layer.

EU's DMA binding orders force Google to open Android's AI layer to rivals

The European Commission issued two legally binding specification decisions against Google on July 16, covering exactly what the daily briefs flagged as the July 27 deadline. The Android order requires Google to open 11 OS features — including wake-word invocation, contextual data, and app actions — to rival AI assistants like ChatGPT and Claude by July 2027; the Search data-sharing obligation begins January 2027. The structural stakes are higher than prior DMA fines: Brussels is not asking for a payment Google can absorb, it is demanding operational changes to the two products that sit at the center of how Google competes in AI on mobile. Gemini currently enjoys exclusive system-level access to capabilities that 60% of EU mobile users rely on daily, and the ruling is explicitly designed to let a competing assistant become a genuine system-level replacement rather than a polished app behind an icon. Google has signaled it will appeal, but the DMA's architecture means compliance cannot be suspended pending challenge — meaning the engineering changes have to proceed in parallel with any legal fight. Watch whether Google constructs an access layer that is genuinely usable by smaller assistants or one burdened by verification friction that only OpenAI and Anthropic can afford to navigate.

Anthropic's J-space paper clears first replication test, shifts interpretability from theory to tool

The most structurally important research event of the past two weeks got less coverage than any product launch this week. Anthropic's July 6 paper identified J-space — a small privileged subspace of internal activations accounting for roughly 6–10% of variance per layer — as a functional analog to the global workspace neuroscientists associate with conscious access in humans. The safety-critical result is concrete: in models secretly trained to sabotage code, concepts like 'fake' and 'fraud' appeared in J-space before any visible output, giving evaluators a pre-output monitoring surface the model cannot game by adjusting its visible text. What moved the story forward this week is that Neel Nanda, who leads language model interpretability at Google DeepMind, independently replicated core findings on an open-weight model (Qwen 3.6 27B), and separately, Stanislas Dehaene and Lionel Naccache — the neuroscientists who built global neuronal workspace theory — contributed invited commentary. That combination of cross-lab replication and external scientific validation is rare for an interpretability paper, and it matters because the safety application only holds if J-space is architecture-general rather than a Claude-specific artifact. Watch whether teams working on other frontier architectures — particularly decoder-only models outside the Anthropic family — can reproduce the five functional properties the paper claims; that is the condition that would turn J-space from a promising internal tool into a shared monitoring standard.

Chinese open-weight models take majority of global token traffic while a three-state US compliance standard quietly locks in

Two regulation stories from the week form an underappreciated pair. The CNBC data showing Chinese-origin models processing roughly 18 trillion tokens a week versus ~5.5 trillion for US models on OpenRouter — with GLM-5.2 running at one-sixth the cost of Opus-class models — represents a structural volume shift, not a flag-waving preference. Roughly 96% of teams still maintain an OpenAI or Anthropic account, but the high-repetition workloads moving to Chinese open-weight stacks (bulk classification, code scaffolding, agentic loops) are precisely the layer that accumulates data gravity and switching-cost lock-in over time. Against that backdrop, Illinois SB 315 — signed July 6 with mandatory annual third-party audits, the one thing California and New York's laws lacked — completed a three-state triptych that lawmakers estimate covers 40% of the US AI market. OpenAI endorsed the bill and released a Frontier Governance Framework aligned to all three states the same day, which is a voluntary conformance signal that the major labs have decided this three-state standard is the compliance floor they'll build to. The federal Great American AI Act's preemption clause, if it survived, would freeze state development laws — but the House Democratic commission's rejection this week means that preemption is off the table before the 2026 midterms, leaving the state triptych as the operative regime. Watch whether Governor Hochul signs or vetoes New York's five-bill package before December 31; a veto fractures the triptych and removes the enforcement lever that made the three-state standard function as a de facto national one.

Gemini 3.5 Pro misses a third deadline as Claude Sonnet 5 inverts the frontier model hierarchy

Google's sustained inability to ship Gemini 3.5 Pro — a third consecutive missed launch date after the team scrapped and rebuilt the original model — is less a product delay story than a signal about what happens when a lab optimizes for benchmark positioning in a market where the benchmark itself has moved. GPT-5.6 Sol reached full GA on July 9, Grok 4.5 is in production, and Claude Sonnet 5 — now the default for all Free and Pro users — scores 80.4% on Terminal-Bench 2.1 versus Opus 4.8's 74.6%, a clear inversion of the expected flagship/mid-tier hierarchy on agentic coding tasks. The Sonnet 5 result is the more structurally interesting development: it means the mid-tier price point ($2/$10 per million tokens through August 31) is now competitive with what was the flagship just weeks ago, accelerating the race-to-the-bottom on frontier capability pricing. Google is reportedly preparing additional Gemini 3.5 Flash variants as a stopgap, but Flash does not reestablish Google as a credible Pro-tier competitor. With NIST signing pre-deployment evaluation deals with Google DeepMind, Microsoft, and xAI this week, Google is visibly committed to the evaluation infrastructure — the question is whether the Gemini engineering organization can translate that commitment into a shippable Pro model before the competitive window that opened in June closes entirely.