Anthropic Bets Its Future on Google TPUs With Multi-Gigawatt Compute Deal
π§ LAUNCH
Anthropic Bets Its Future on Google TPUs With Multi-Gigawatt Compute Deal
Anthropic has signed a deal with Google and Broadcom for multiple gigawatts of next-generation TPU capacity, coming online starting in 2027 to train and serve frontier Claude models. This is the largest known compute commitment by any non-Google company β and it's a deliberate bet against the NVIDIA-only training paradigm that every other frontier lab relies on. The strategic signal is loud: Anthropic is locking in custom silicon at infrastructure scale while competitors fight over the same GPU allocation pool. Watch for Claude model quality jumps as this capacity comes online. (7,442 likes | 483 RTs) Read more β
NVIDIA Officially Quantizes Gemma 4 31B to Run on a Single Consumer GPU
NVIDIA dropped an official NVFP4 quantization of Gemma 4 31B on Hugging Face β 4x smaller weights while maintaining frontier-level quality. This isn't a community quant with unknown quality trade-offs; it's a vendor-blessed compression that makes a 31B-parameter model runnable on consumer hardware. If you've been waiting to ditch API calls for local inference, this is your on-ramp. (3,596 likes | 374 RTs) Read more β
Gemma 4 E2B lands as Google's any-to-any multimodal variant β handling multiple modalities in both input and output directions. Already racking up 23K downloads with 134 likes, it adds a new capability tier to the Gemma family that positions it directly against proprietary multimodal competitors. (134 likes | 23.1K downloads) Read more β
π§ TOOL
X/Twitter Releases Official MCP Server for AI Agent Access
X just shipped an official MCP server, meaning your AI agent can now search tweets, read bookmarks, and interact with X programmatically through the Model Context Protocol. This is a major validation of MCP as the emerging standard for AI-tool integration β when a platform the size of X builds a first-party implementation, the protocol war is effectively over. Clone it, connect it, and give your agent the firehose. For more on setting up MCP integrations, see our MCP setup guide. (820 likes | 68 RTs) Read more β
Claude gets real-time financial data via a new Financial Datasets MCP Server β live stock prices, balance sheets, cash flow statements, and breaking company news. This eliminates the hallucinated-financials problem that's been the top complaint in finance use cases. If you're using Claude for any financial analysis, set this up today. (486 likes | 65 RTs) Read more β
Gradio Server decouples Gradio's backend from its frontend β you can now use React, Vue, or any custom framework as the UI while keeping Gradio's ML serving, auth, and queue infrastructure. Best of both worlds for anyone who loves Gradio's backend but hates its default look. Read more β
OpenAI hosts a live deep-dive on Codex workflows with the team that built it β covering everything from feature exploration to team shipping patterns. Practical, not theoretical. If you're evaluating Codex for your team, check out our Codex vs Claude Code comparison for context. (196 likes | 12 RTs) Read more β
π TECHNIQUE
Fine-tune Gemma 4 in a free Colab notebook: Unsloth Studio now supports Gemma 4 and 500+ other open-source models β no setup, no GPU costs, just click run. The fine-tuning barrier just dropped to literally zero dollars. (656 likes | 86 RTs) Read more β
Running Gemma 4 locally with LM Studio and Claude Code: A practical step-by-step guide combining LM Studio's new headless CLI with Claude Code for local inference. Includes real benchmarks so you know exactly what to expect before committing. (157 likes | 40 RTs) Read more β
π¬ RESEARCH
Apple Research Shows AI Models Fail at Grade-School Arithmetic
Apple just published research demonstrating that current AI models β all of them β fail at basic arithmetic a ten-year-old could handle. Amplified by Yann LeCun to 2.4K likes, this is a sobering reality check for anyone deploying math-dependent AI features in production. If your product relies on AI doing calculations, you need a calculator fallback, not a prayer. (2,429 likes | 639 RTs) Read more β
OpenAI launches a Safety Fellowship funding independent researchers to work on alignment and safety outside OpenAI's own team. This directly addresses the criticism that safety research is too centralized inside the labs that build the models β a meaningful structural change, not just a PR move. (1,679 likes | 181 RTs) Read more β
Nature publishes a paper on fully automating AI research end-to-end β from hypothesis generation through experimentation to paper writing. The meta-recursive angle is hard to ignore: AI accelerating AI research is no longer speculative, it's published in Nature. (271 likes | 56 RTs) Read more β
π‘ INSIGHT
Anthropic's Revenue Triples to $30B Run-Rate in Four Months
Anthropic's annualized revenue surged from $9B to $30B in roughly four months β the fastest revenue growth in AI history. This isn't hype-cycle growth; it's Claude converting to real enterprise spend at a rate that validates the entire "safety-first premium model" thesis. For context, OpenAI took over a year to make a comparable jump. (1,506 likes | 117 RTs) Read more β
Meta is preparing its first LLM built under Alexandr Wang's leadership, targeting specific consumer strengths rather than trying to beat frontier labs across every benchmark. Plans to eventually open-source versions. A different strategic posture from the "Llama catches up to GPT" playbook. (348 likes | 16 RTs) Read more β
The Axios supply chain attack started with social engineering a maintainer. Simon Willison flags the attack vector as a warning: increasingly sophisticated social manipulation targeting individual open-source developers. If you maintain any popular package, review your contributor access controls now. (437 likes | 73 RTs) Read more β
Karpathy argues AI could flip the surveillance relationship β citizens making governments legible instead of the reverse. With 4.9K likes, this reframes AI's civic potential beyond the usual productivity narrative. What government data would you want an AI to parse for you? (4,916 likes | 600 RTs) Read more β
ποΈ BUILD
Eight years of wanting, three months of building with AI: A developer ships SyntaqLite β a project deferred for nearly a decade, finally built in three months with AI assistance. At 573 HN points, it's the most concrete case study yet of how AI is unlocking ambitious side projects for experienced engineers who previously lacked the time. Read more β
Hippo: A hippocampus-inspired memory system for AI agents that models consolidation, forgetting, and retrieval patterns from neuroscience. Goes well beyond naive vector stores β if you're building agents with long-term memory needs, this is worth a look. (30 likes | 12 RTs) Read more β
π MODEL LITERACY
Model Quantization (NVFP4): When NVIDIA quantized Gemma 4 31B to NVFP4 format, they shrank the model's weights by 4x β turning a model that needed a multi-GPU server into one that runs on a single consumer GPU. Quantization works by reducing the numerical precision of each weight (from 16-bit or 32-bit floating point down to 4-bit), trading tiny amounts of accuracy for massive memory savings. The reason official vendor quantizations like NVFP4 matter more than community ones: NVIDIA calibrates the compression against known benchmarks and optimizes for their hardware's specific number formats, resulting in less quality loss than a generic quantization recipe. If a model's weights fit in your GPU's VRAM, you can run it locally β and understanding quantization formats is how you know which models will actually fit.
β‘ QUICK LINKS
- Anthropic's TPU deal tweet: The social proof β 7.4K likes on the compute announcement. Link
- AI singer 'Eddie Dalton': Holds 11 simultaneous iTunes chart spots despite not being human. The music industry's AI reckoning is here. Link
- Mollick on Gemma 4: Impressive on-device, but can't handle real agentic workflows β judgment and self-correction gaps remain. (386 likes) Link
- Claude Code "unusable for complex tasks": GitHub issue hits 709 upvotes β read it for the failure modes and workarounds. Link
π― PICK OF THE DAY
The first multi-gigawatt TPU commitment by a non-Google company just reshaped the AI infrastructure map. Anthropic's deal with Google and Broadcom isn't just about getting more compute β it's a declaration that the frontier lab race has become an infrastructure race. While every other major AI company fights over the same NVIDIA GPU allocations, Anthropic is building an alternative supply chain from scratch. The timing is telling: with revenue tripling to a $30B run rate, Anthropic has the cash to make a bet this size, and the conviction that custom TPU silicon can match or beat NVIDIA for training frontier models. If this works, it forces every frontier lab to rethink their hardware strategy β do you keep competing for the same shrinking pool of H200s, or do you find your own silicon partner? The compute supply chain just went from a single bottleneck to a genuine market, and that's good for everyone building on top of these models. Read more β
Until next time βοΈ