One-Line Summary

Export restrictions and tight compute supply are reshaping who gets access to top AI models — prompting regional launches in Asia, capacity rationing at Big Tech, and massive GPU procurement deals.

Big Tech

Google caps Meta’s Gemini usage amid capacity strain

Google is limiting how much of its Gemini AI Meta can use after telling the company around March that it could not meet the full capacity Meta sought, Reuters reports, citing the Financial Times. The shortfall reportedly disrupted or delayed some of Meta’s internal AI projects. ¹

Several other Google clients also face constraints, and Meta has asked staff to be more efficient with “AI tokens,” the units that measure usage. Google Cloud revenue grows to $20 billion in the quarter ended March, but CEO Sundar Pichai says compute power limits held growth back and expanded the backlog. ¹

Alphabet leans on TPUs to compete for AI compute

Google’s tensor processing units (TPUs) are in‑house AI chips that power Gemini and are available through Google Cloud; customers including Anthropic rent access and, in some cases, can buy TPUs for their own data centers, CNBC writes. The chips position Alphabet as a major force in AI infrastructure alongside Nvidia’s GPUs. ²

CNBC, citing FactSet, notes Wall Street projects Google Cloud revenue to climb about 64% in 2026 to $96 billion, while supply constraints — from memory to manufacturing — remain a risk for the entire compute stack. ²

Meta tests prediction app ‘Arena’ and eyes Polymarket/Kalshi tie‑ups

Meta is developing a points‑based prediction market app called Arena and, at CEO Mark Zuckerberg’s request, is exploring partnerships with Polymarket and Kalshi, Reuters reports, citing the New York Times. Meta and Polymarket decline to comment; Kalshi does not respond. ³

The report says Arena targets 18‑ to 34‑year‑olds, aims for at least 100 million monthly “predictors,” is being tested internally, and may not be released; Meta plans to eventually integrate parts into Facebook and Messenger. ³

Industry & Biz

Asian players move into Mythos gap as U.S. export curbs bite

Two Asian companies roll out offerings positioned against Anthropic’s restricted models: Tokyo‑based Sakana AI launches Fugu, which can coordinate access to other models via APIs, and China’s 360 unveils Tulongfeng to find software vulnerabilities (alongside Yitianzhen for automated cyber defense), TechCrunch reports. The momentum follows U.S. orders limiting Anthropic’s global access to Mythos and Fable. ⁴

Sakana says Fugu targets Japanese businesses and government agencies seeking to reduce exposure to tightening export controls, while insisting U.S. models remain important to Asia. The company advertises “frontier capability without the risk of export controls” and frames Fugu as an orchestration model for agents. ⁴

TechCrunch adds that 360’s founder Zhou Hongyi calls vulnerability‑finding AI a strategic national asset and warns of “one‑way transparency” if only some actors access advanced detection tools. ⁴

Anthropic’s revenue run‑rate crosses $47 billion in May 2026, TechCrunch writes, but with Mythos and Fable restricted abroad, local alternatives tuned for language and culture begin filling gaps in enterprise demand. ⁴

Firmus inks Nvidia deal for 170,000 GPUs and shared cloud revenue

Australian AI infrastructure firm Firmus Technologies signs a strategic partnership with Nvidia to buy Nvidia infrastructure and sell Nvidia‑powered cloud services, with Nvidia earning product revenue and a share of cloud revenue, Reuters reports. Firmus positions the pact as lowering barriers for “AI Native” up‑and‑comers. ⁵

The deal delivers 170,000 GPUs between the first quarter of 2027 and early 2028 in Batam, Indonesia, and Firmus expects up to $30 billion in revenue over the first six years based on customer commitments, aiming to give smaller AI firms more affordable access to compute, co‑CEO Tim Rosenfield says. ⁵

Community Pulse

Hacker News (265↑) — Skepticism about “Mythos‑like” performance claims mixes with concern that export limits could accelerate a China shift. ⁶

"Well, goodbye to US tech and welcome China then I guess, nobody will wait and you can't put the genie back in the bottle." — Hacker News ⁶

Hacker News (134↑) — Users find Gemini useful for some tasks but report hallucinations and frequent capacity‑related downtime. ⁷

"Not really, especially recent gemini's tend to hallucinate unbelievably much especially with visual input. And their safety tuning is neither effective nor precise on edge models." — Hacker News ⁷

"We use Gemini for some specific tasks. It is often unavailable due to capacity limits or other downtime. It's probably the best multimodal model I've worked with (if somebody knows a better one for audio analysis, please let me know!)" — Hacker News ⁷

Hacker News (105↑) — Many see a prediction app as lucrative given behavior on social platforms, while questioning the morality and tone of such a product. ⁸

"Is it just me or is this really not that crazy of an idea? Polls are already a thing, and people love giving their opinion on things. Combining social media and gambling while maybe not morally great. Seems like has potential to make a shitton of cash. Betting is all over the place right now." — Hacker News ⁸

"Facebook is the 21st century version of the tabloid newspaper and daytime Jerry Springer. Highly lucrative, but bottom of the barrel stuff. Launching a prediction market is exactly the type of thing I would expect from them." — Hacker News ⁸

What This Means for You

Export rules and capacity ceilings can change who has access to a model overnight. If your workflows depend on a single provider, plan for continuity so an export decision or quota doesn’t halt campaigns, data analysis, or customer support automations. Dual‑source where possible and keep prompts and integrations portable. ⁴

Expect usage budgeting to become a daily reality. Meta’s internal push to conserve “AI tokens” mirrors what many teams face as providers meter access — set per‑project limits, track cost per task, and prioritize high‑ROI use cases first. ¹

Regional options are getting more credible for local‑language and compliance‑sensitive work. For Japan‑facing products, a model tuned to Japanese language and culture could reduce review cycles or improve response quality, while serving as a hedge against cross‑border restrictions. ⁴

If you build engagement products, a points‑based prediction experience may surface in major social apps. That could offer new acquisition or retention tactics, but it also raises moderation, integrity, and legal review needs before any experiment. ³

Action Items

1. Draft a one‑page LLM continuity plan: List your primary and backup model providers, how you would switch in a day, and what data/contracts are needed to do it safely.
2. Set token budgets and alerts: In your AI platform console, create per‑team usage quotas and alerts so capacity limits don’t surprise a live campaign.
3. Request information on regional models: Ask vendors for demos or eval access to Japan‑ or APAC‑tuned models and to any “orchestration” features that route across multiple models.
4. Harden AI security workflows: With new AI‑assisted vuln‑discovery tools emerging, document which tools are approved, how outputs are verified, and who signs off on remediation.
5. Prototype a points‑only prediction poll: If relevant to your product, run a limited, non‑monetary prediction poll with moderation and legal review to test engagement safely.