Digital Colliers Daily Briefing — June 28, 2026
Compute scarcity, export politics, and the talent war for the next consumer device converged today to illustrate how unevenly power is distributing across the AI stack. The Financial Times reports Google rationed Gemini capacity to Meta as far back as March, slowing some of Meta's internal AI work. The fallout from Washington's Mythos/Fable export ban is producing rival Asian frontier models and a Chinese gray market for Claude. And Apple's Vision Pro chief, who also oversaw the company's forthcoming smart glasses, is decamping for OpenAI's Jony Ive–led hardware effort.
1. Google Rations Gemini Capacity to Meta, Slowing Internal AI Roadmap
What happened. According to the Financial Times, Google told Meta around March that it could not supply all the Gemini capacity Meta had asked to purchase. The shortfall disrupted and delayed several internal Meta AI projects that had been built around access to Google's frontier models through Google Cloud.
Why it matters. The episode is a concrete data point in a debate that has so far been mostly anecdotal: frontier compute and frontier-model access are now scarce enough that a hyperscaler will throttle a paying customer — even one of the world's largest technology companies. It also surfaces an awkward duality in the cloud-AI market. Google sells Gemini as a commercial API while simultaneously competing with Meta in consumer AI, advertising, and developer platforms. The FT framing — that surging appetite has turned computing power into the industry's scarcest commodity — implies allocation decisions are increasingly strategic rather than purely commercial.
Who is affected. Meta's product teams most directly: any roadmap that assumed elastic Gemini access has slipped. More broadly, the news will harden the conviction inside every large AI buyer that single-vendor dependence on a rival's models is a structural risk, reinforcing multi-model strategies and accelerating in-house training investments. For Google Cloud, the disclosure complicates its pitch to enterprise customers who need to trust that capacity commitments will hold.
What to watch next. Whether Meta publicly reframes its Llama roadmap or accelerates custom-silicon and training capacity spending in response; whether other Gemini customers — particularly those competing with Google's own product lines — report similar throttling; and whether Microsoft and AWS use the moment to court Meta workloads.
Sources:
2. Mythos Export Ban Spawns Sakana's Fugu, 360's Tulongfeng, and a Claude Black Market
What happened. Two weeks into the U.S. order barring Anthropic from offering its Mythos and Fable 5 cybersecurity-tuned models to non-Americans, two Asian competitors moved into the gap. Tokyo-based Sakana AI launched Fugu, an orchestration-oriented frontier model its team says "stands shoulder-to-shoulder" with Fable 5 and Mythos Preview. Chinese cybersecurity firm 360 unveiled Tulongfeng for vulnerability discovery and Yitianzhen for incident response; founder Zhou Hongyi framed vulnerability-finding AI as a national strategic asset, warning of "one-way transparency," according to Reuters reporting cited by TechCrunch. Wired separately documents a thriving Chinese underground for Claude access, including "transfer station" sites that buy API tokens abroad and resell them via proxies and Telegram-sourced identities. The Wall Street Journal reports researchers now find Z.ai's GLM-5.2 matches the latest U.S. models at security-bug discovery. And Politico reports AI executives are quietly seeking regulatory clarity from the Trump administration but fear retaliation if they push too hard — an unpredictability underscored Friday when the administration partially rescinded the Anthropic ban.
Why it matters. The export regime is producing the outcome critics predicted: rather than denying China and the broader Asian market access to frontier cyber-capable AI, it is catalyzing local substitutes and a smuggling economy. Sakana co-founder Ren Ito has used a Project Syndicate op-ed and remarks at the G7 in Evian to argue that Washington's "first priority should be to preserve access" for allies. Co-founder David Ha is using the moment to pitch orchestration models — coordinating many providers — as a structural hedge against single-vendor dependence. A widely circulated Hacker News analysis from Cephalosec (Post-Mythos Cybersecurity: Keep calm and carry on) argues Mythos's gains over Opus 4.6 and GPT-5.4 are real but more linear than the marketing implied, noting that the BSD finding cited in Project Glasswing required roughly $20,000 in inference per discovery and that open models like DeepSeek, Gemma 4, and Qwen 3.6 find about half the vulnerabilities Mythos surfaces in benchmarks. The piece also flags OpenAI's new Sol/Terra/Luna family as a defensively tuned competitor now subject to similar U.S. vetting.
Who is affected. Anthropic, which reported a $47 billion run rate in May, faces an erosion of Asian enterprise demand it cannot easily measure or recover. Japanese government and enterprise buyers — Sakana's stated target — gain a domestically aligned alternative. Chinese defenders and offensive teams alike gain locally produced tooling at exactly the moment Washington wanted to widen the gap. U.S. cybersecurity firms with Glasswing-tier access enjoy temporary artificial scarcity that they monetize at premium, a dynamic the Cephalosec post calls out directly.
What to watch next. Whether the partial rescission noted by Politico expands into a broader rollback; independent benchmarks comparing Fugu and Tulongfeng to Fable 5; Anthropic's enforcement posture against the token-resale market Wired describes; and whether the WSJ's GLM-5.2 findings prompt Commerce to revisit how it treats Chinese open-weights models.
Sources:
- Asian AI startups launch Mythos-like models as Anthropic's export ban drags on — TechCrunch AI
- AI executives and lobbyists say they are seeking regulatory clarity from the Trump administration but are wary of pressing for answers, fearing retaliation (Politico) — Techmeme
- A look at a thriving underground economy for Claude access in China, including "transfer station" sites that buy API tokens abroad and distribute them to users (Wired) — Techmeme
- Tokyo-based Sakana AI's Fugu and China-based 360's cybersecurity model Tulongfeng claim to rival Anthropic's banned Mythos and Fable 5 amid the US export ban (Kate Park/TechCrunch) — Techmeme
- Post-Mythos Cybersecurity: Keep calm and carry on — Hacker News
- Researchers say Z.ai's GLM-5.2 matches latest US models at finding security bugs, as critics question the US' lax approach in restricting Chinese open models (Wall Street Journal) — Techmeme
3. Paul Meade, Apple's Vision Pro and Smart Glasses Lead, Joins OpenAI Hardware
What happened. Paul Meade, the Apple vice president running the Vision Pro headset and also leading development of Apple's AI-powered smart glasses scheduled for next year, is leaving to join OpenAI's hardware team, according to Bloomberg's Mark Gurman as reported by TechCrunch. Gurman attributes the departure in part to John Ternus's imminent elevation to Apple CEO and a hardware-engineering reorganization that has left several VPs feeling demoted.
Why it matters. OpenAI's hardware effort, led by former Apple chief design officer Jony Ive, has been visibly struggling — reports last fall described difficulty nailing down the product's form factor for what Sam Altman has described as a calmer alternative to the iPhone. Recruiting the executive who shipped Vision Pro and was building Apple's Meta Ray-Ban competitor brings exactly the operational and supply-chain expertise that pure design talent does not. It also lands at a delicate moment for Apple: Vision Pro underperformed commercially, the smart glasses program was meant to be the correction, and the person running both is now walking out the door during a CEO transition.
Who is affected. Apple's wearables roadmap loses institutional memory at the program-leadership level just as it tries to ship a device aimed squarely at Meta's Ray-Ban line. OpenAI gains a credible hardware operator to pair with Ive. Meta, which has so far enjoyed a largely uncontested run in AI-native eyewear, now faces two better-resourced challengers on convergent timelines.
What to watch next. Whether additional Apple hardware VPs follow Meade out under the Ternus reorganization; any shift in Apple's announced 2027 smart glasses timeline; and concrete signals — patents, supplier leaks, FCC filings — about what form the OpenAI/Ive device will take.
Sources:
Today's three stories trace the same underlying scarcity from different angles: compute, models, and the people who turn them into products. Google's rationing and Washington's export controls both show that frontier AI capacity is now allocated politically as much as commercially, while Meade's move underscores that the talent capable of translating models into shipping hardware is just as constrained. The companies positioned to absorb that scarcity — through in-house silicon, orchestration strategies, or aggressive hiring — are the ones setting the terms for the next eighteen months.