OpenAI keeps pulling Apple's best hardware people across the table, and the device race is turning into a talent race. The cost of running AI is finally biting, with open models like GLM-5.2 and DeepSeek pushing teams to rethink their bills. Europe, meanwhile, is so rattled by US export controls that Austria wants to host Anthropic itself. Here is what matters today, and what to do about it.

Bloomberg reported on June 26 that Paul Meade, who ran Apple's Vision Pro hardware engineering for seven years and led its smart-glasses program, is leaving to head a new OpenAI hardware unit. He joins former Apple figures including Jony Ive and Evans Hankey, and reports say OpenAI has recruited more than 40 Apple hardware engineers for a screenless, voice-first device targeted no earlier than early 2027. His exit follows a reshuffle under incoming Apple CEO John Ternus.

Why it matters: The AI race is now also a race for the people who can turn models into shippable hardware, and Apple is losing the talent behind its most ambitious products just as its smart glasses slip to late 2027.

What to do:

Source: TechCrunch

GLM-5.2, Z.ai's MIT-licensed open model (about 753 billion parameters, roughly 40 billion active per token), now scores 62.1 on SWE-bench Pro, ahead of GPT-5.5 at 58.6, at roughly one-sixth the hosted cost, with fresh benchmarks still landing. The pressure is showing up on invoices: CNBC reported on June 26 that AI startup Lindy moved 100% of its traffic to DeepSeek to cut costs, and DeepSeek's permanent price cut ($0.44 and $0.87 per million input and output tokens) undercuts GPT-5.5 by more than 5x on input. Enterprises that once spent freely are now scrutinising token spend as both OpenAI and Anthropic eye IPOs.

Why it matters: The old trade-off, cheap open weights versus production-grade reasoning, is breaking down, which pressures the pricing power of closed labs and rewards teams that route each task to the cheapest model that can handle it.

What to do:

Source: CNBC

After a US Commerce Department directive on June 12 barred foreign nationals from Anthropic's most advanced models, Fable 5 and Mythos 5, forcing a global shutdown of both, Austria's State Secretary for Digitalization wrote to the EU urging Europe to court Anthropic to set up locally. Reuters and Bloomberg reported the letter on June 28, which frames frontier-AI access as European sovereignty: a single market of 450 million people, it says, was cut off "at the stroke of a pen." A partial rollback let vetted US institutions back onto Mythos on June 26, but non-US access stayed blocked.

Why it matters: Governments are now competing to host AI labs as US rules tighten who can use the most capable models, a new geopolitical risk for any business that depends on one jurisdiction's AI.

What to do:

Source: Bloomberg

A solo developer populated a private World of Warcraft server with around 1,800 AI-driven bots, using the AzerothCore emulator for gameplay and the DeepSeek API for chat, for under £10 a month. The bots quest, level and chatter among themselves, making a near-empty realm feel alive in a "proof of concept" shared on Reddit and amplified by a former Blizzard developer.

Why it's interesting: It shows AI being used to fake social presence rather than solve hard problems, and reignites the "dead internet" debate about single-player MMOs full of synthetic players.

Key takeaway: AI can keep ageing online worlds feeling populated long after the humans leave.

Source: GameSpot

A developer found OpenAI's Codex CLI quietly writing about 37TB of logs to an SSD over 21 days, an annualised rate near 640TB that could exhaust a consumer drive's endurance in under a year. The culprit was a TRACE-level SQLite logger that ignored the usual log setting. OpenAI merged fixes by June 23, and the issue is now closed, cutting roughly 85% of the writes.

Why it's interesting: It is a reminder that fast-moving AI tooling can quietly wear down real hardware while looking calm in any disk-space view.

Key takeaway: Update Codex, and keep an eye on what your AI dev tools write to disk.

Source: GitHub

Getty Images announced a multi-year display deal on June 21 that surfaces its licensed photo and editorial library, over 400 million assets, inside ChatGPT's search answers, and Getty's stock spiked as much as 145% intraday. The deal is display-only and grants no training rights, a striking reversal for a company that sued Stability AI over AI imagery in 2023.

Why it's interesting: A firm that once fought generative AI in court is now one of its content partners, and the market rewarded the truce instantly.

Key takeaway: Licensing, not litigation, is becoming the path for media companies dealing with AI.

Source: Getty Images

Context: With open models closing the gap on cost, the fastest win is routing each task to the cheapest model that can do it well, instead of paying frontier prices for everything.

Prompt: I run [team or product] and my current AI stack is [models and tools you use] costing roughly [monthly spend]. List my main AI tasks (for example drafting, summarising, coding, research, support). For each, recommend whether to use a premium, mid-tier, or cheap or open model like GLM-5.2 or DeepSeek, and explain the quality risk of moving it down a tier. Then give me a migration order ranked by biggest cost saving for least risk.

Do not: Do not move customer-facing or high-stakes reasoning tasks to a cheaper model without a side-by-side quality test first.

If/Then: If a task involves legal, financial, or safety-critical judgement, then keep it on the premium model and flag it for human review.

Example: A 12-person agency, ACME Studio, ran this and moved bulk transcription and first-draft copy to an open model, kept client strategy on a premium model, and cut its monthly AI bill by about 40% with no drop in client-facing quality.

Anthropic launched Claude Tag on June 23, letting Enterprise and Team customers add @Claude as a standing member of a Slack channel that builds memory over time, works asynchronously across hours or days, and in ambient mode revives stale threads on its own. It runs on Claude Opus 4.8, every action is logged, and admins control its tool and data access. Anthropic says the internal version already writes about 65% of its product team's code; the old Claude in Slack app retires on August 3.

Why it's important: This shifts AI from a tool you query to a coworker embedded in the workflow, deepening both its usefulness and the vendor lock-in.

Business takeaway: Pilot it in one channel with strict admin scopes and a token cap before any wider rollout.

Source: TechCrunch

A new Exponential View report pegs genuine end-customer generative-AI revenue at $110 billion over the trailing 12 months, with the run rate now past $175 billion, in what it calls the first bottom-up, deduplicated measure of the market. It estimates AI is scaling about 3x faster than the mobile and internet waves; a fresh $1 billion of revenue now arrives in under two days, versus 180 days in 2023; and AI has nudged a flat US power sector back to growth at roughly 9 TWh a month.

Why it's important: It reframes AI as a fast-compounding market whose biggest economic effects may still be ahead, even as spend only just covers the hardware depreciation behind it.

Business takeaway: Falling token prices keep expanding usage, so budget for more AI consumption, not less, as you scale.

Source: Exponential View

OpenAI and Broadcom unveiled Jalapeño on June 24, OpenAI's first custom-designed inference chip, aimed at serving its models roughly 50% cheaper than the Nvidia GPUs it has leaned on. It is not commercially available; Broadcom expects small prototype deployment by the end of 2026, with the partners targeting accelerators at 10-gigawatt scale through 2029. OpenAI spent an estimated $14 billion serving ChatGPT on third-party GPUs in 2025, so the cost lever is large.

Why it's important: Owning the inference silicon is how OpenAI closes its structural cost gap with rivals that already run custom chips.

Business takeaway: Cheaper inference at scale should keep pushing AI prices down over the next two years.

Source: Build Fast with AI

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas