Google told Meta it could not have all the Gemini compute it wanted. That is not a rumor about a future problem. It happened in March, and the consequences are still rippling through Meta’s AI projects months later. The Financial Times reported it Sunday, and CNBC and Bloomberg carried it across the day.
For anyone building on frontier APIs, the story is not really about Google and Meta. It is about what happens when the biggest buyers in the world cannot get enough inference capacity from the biggest providers. If Meta gets rationed, everyone downstream is on borrowed time.
The big signal
CNBC reported, citing the Financial Times, that Google put limits on Meta’s use of its Gemini AI models after Meta sought more computing capacity than Google could provide. Google told Meta around March it could not meet the full Gemini capacity Meta had sought to purchase. The shortfall disrupted and delayed some of Meta’s internal AI projects. Several other Google clients were also affected, though to a lesser extent.
Meta has encouraged staff to be more efficient with AI tokens, the units that measure AI usage, the FT report said. That detail matters. When a company the size of Meta has to tell engineers to ration their AI token usage, you are looking at a supply constraint that has moved from the chip layer to the application layer.
This is not a prediction. Google Cloud revenue grew to $20 billion in the first quarter, but CEO Sundar Pichai said computing power constraints prevented even higher growth and contributed to the cloud unit’s backlog nearly doubling quarter on quarter. The demand is real. The supply is not keeping up. And the biggest AI companies are now competing for inference capacity the way they once competed for GPUs.
For practical builders, the implication is direct. If you are building a website assistant, an internal agent, or any product that depends on a frontier model API, your cost and availability assumptions need stress-testing. The model may be available today, but the compute behind it is not guaranteed. Multi-model routing, caching, and fallback to smaller or open-weight models are no longer just cost optimizations. They are continuity planning.
China’s Z.ai and Qihoo 360 claim Mythos-level cybersecurity AI
While compute scarcity dominates the supply side, the capability race on the other side is getting sharper. The Decoder reported that Chinese cybersecurity firm Qihoo 360 has built two AI tools designed to rival Anthropic’s Mythos. The tools, called Tu Long Feng and Yi Tian Zhen, automate vulnerability discovery and cyber defense respectively. Tu Long Feng has already flagged 3,432 vulnerabilities, according to Qihoo 360 founder Zhou Hongyi.
Separately, Z.ai founder and Tsinghua professor Jie Tang estimated that a Chinese “Mythos-class” model would arrive before Q1 2027. Z.ai recently shipped GLM-5.2, an open-weight model that has drawn attention for reportedly approaching Claude-level performance on cybersecurity tasks. The Verge, WSJ, and The Information all carried versions of the story.
Zhou framed the race in explicitly strategic terms, calling Mythos the equivalent of “cyber-nuclear weapons of the AI era” and arguing that China needs its own deterrent. He put the gap between top Chinese models and the most capable Western ones at 20 to 30 percent, which is closer than many in Washington assumed even six months ago.
For the AI community, the signal is not that China has caught up. The signal is that the export control strategy designed to slow China down may be accelerating the development of domestic alternatives. When the US restricts access to frontier models, it does not eliminate demand. It redirects it. And some of that redirected demand is producing capable open-weight models that the rest of the world can use.
Open-source watch
DeepSeek DSpark. Tech Times reported that DeepSeek released DSpark, a speculative decoding framework that makes DeepSeek V4 up to 85 percent faster. Speculative decoding is not a new idea, but DeepSeek’s implementation could meaningfully reduce inference costs for teams running open-weight models. For anyone serving models locally or on their own infrastructure, faster inference directly translates to lower operating costs and better user experience.
GLM-5.2 open weights. Z.ai’s GLM-5.2 continues to generate attention as an open-weight model that reportedly approaches Claude on cybersecurity tasks. Indian Express and multiple outlets covered it as a model “turning heads in Silicon Valley.” Whether it truly matches Claude on production workloads is unverified, but the pattern matters: Chinese labs are shipping capable open-weight models at a pace that Western policy was not designed for.
Liquid AI LFM2.5-230M. MarkTechPost reported that Liquid AI shipped a 230-million-parameter model with support for llama.cpp, MLX, vLLM, SGLang, and ONNX. That is a model small enough to run on-device across a wide range of hardware. For privacy-first applications and local AI deployment, sub-300M models with broad runtime support are exactly what makes on-device agents practical.
Citi: open-source benefits from frontier restrictions. Investing.com reported that Citi analysts see open-source AI models benefiting directly from restrictions on frontier systems. The logic is straightforward: when the best closed models become harder to access, the gap between closed and open narrows in practice even if it widens in theory.
HP and OpenAI go enterprise-wide on the Frontier platform
Thurrott.com reported that HP announced a strategic partnership with OpenAI to integrate OpenAI’s Frontier platform across HP’s customer-facing and internal operations. HP plans to deploy AI-driven solutions across customer and partner experiences, telemetry insights through its Workforce Experience Platform, employee productivity, and software development.
OpenAI chief revenue officer Denise Dresser framed it as “AI becoming an operating layer, connected to the systems and workflows where work already happens.” HP also hinted at “agentic AI devices” optimized for always-on inference, which suggests the hardware side is starting to catch up to the agentic software vision.
This is the enterprise AI adoption story that matters for practical builders. HP is not experimenting. It is wiring AI into customer service, partner channels, and internal productivity simultaneously. That is the pattern that small and mid-sized businesses should watch: the move from pilot to operating layer. The companies doing this well are not deploying a chatbot. They are connecting AI to the systems where work already happens and giving it a defined role in the workflow.
What builders should take from this
Three things stand out from the last 24 hours.
- Compute is the new bottleneck, not models. Google rationing Meta’s Gemini access is a structural signal. If you depend on a single frontier API, you need a fallback plan. Multi-model routing, open-weight alternatives, and caching are now operational requirements, not optimizations.
- Export controls are accelerating open-weight development. China’s Z.ai and Qihoo 360 are building Mythos-class alternatives because they cannot access Mythos. The result is more capable open-weight models entering the global market. For builders who care about model sovereignty and local deployment, this is an unexpected upside of restrictions.
- Enterprise AI is moving from pilot to operating layer. HP’s OpenAI partnership shows the pattern. The companies getting value are connecting AI to real workflows, not running isolated demos. Whether you are a small business or a large enterprise, the question is no longer “should we use AI?” but “which workflows should AI own, and how do we make that reliable?”
The practical takeaway
If you are building anything on top of frontier AI, the last 24 hours reinforced one uncomfortable truth: the most capable model is not automatically the safest foundation. Access can be rationed. Capacity can run out. Policy can change overnight. The teams that thrive will be the ones who treat model choice as a supply chain problem, not a benchmark problem.
That means knowing which models you can run locally, which open-weight alternatives are credible fallbacks, and where your product breaks if a single provider cuts your capacity. It also means watching the open-weight space closely, because the models that can run on your own hardware are the ones that cannot be rationed away from you.
The frontier is still where the best capability lives. But the infrastructure around it is getting fragile, and the alternatives are getting better faster than anyone expected.
Leave a Reply