Anthropic just made a staggering $35 billion bet on Claude — here's why it needs so much power

The Hook

Our machine tracks sources around the clock — measuring where the industry converges, not guessing. Today the signal is loud: Anthropic just raised $35 billion, and the real story isn't the money — it's how much silicon that buys, and why frontier AI requires that much raw compute power. We strip the hype, surface the hardware reality, and hand you everything you need before your first meeting.

The Cold Open

Somewhere right now, a data center the size of several city blocks is humming at full capacity. Cooling towers push heat into night air. Inside, row after row of Nvidia H200 GPUs — each one an extraordinarily expensive piece of hardware — burn through floating-point operations at a rate no human brain can intuitively grasp. This is what training a frontier AI model actually looks like. And this morning, Anthropic announced it needs a great deal more of it. The question isn't whether it can build Claude. It's whether the world's semiconductor supply chains can deliver fast enough to spend $35 billion on.

The Signal

1. Anthropic's $35 Billion Compute Bet

Anthropic closed a landmark funding round, underscoring the extraordinary capital now flowing into frontier AI. The 'why it needs so much power' framing is exactly right. Training a frontier model like Claude 4 requires clusters of tens of thousands of Nvidia H200 or Blackwell B200 GPUs running continuously for months. A large GPU cluster carries enormous procurement costs, with substantial ongoing power and cooling expenses on top of hardware. At $35 billion, Anthropic isn't just buying compute for today — it's reserving capacity in a global GPU supply chain that Nvidia, TSMC, and data-center operators are all scrambling to expand. The strategic read: whoever locks in compute commitments in 2026 controls model capability in 2028. Anthropic just made the largest single lock-in bet in the industry's history.

2. ChatGPT Ads Hit $1 Billion Annualized Run Rate

OpenAI's advertising business has crossed $1 billion in annualized revenue — and most users haven't noticed the ads creeping in. The hardware angle is subtle but real: serving ads at ChatGPT scale means running inference on hundreds of thousands of concurrent sessions, each requiring GPU time to generate personalized responses. The marginal cost of every ad impression is denominated in GPU-seconds. As the ads business grows, OpenAI must either expand its compute fleet or get more efficient per inference pass — and both paths run straight through Nvidia's order books. The deeper strategic signal: OpenAI is now a media business, not just an AI lab, and the infrastructure bill for that media business is measured in accelerator hardware and power contracts.

3. AIR Raises $50M for Agent Supply-Chain Security

AIR just closed $50 million to vet the skills, plug-ins, and add-ons that AI agents consume at runtime — naming a new category: agent supply-chain security. From a hardware and infrastructure perspective, every unvetted agent skill that runs in an enterprise environment is not just a security risk — it's a compute liability. Malicious or poorly designed skills can trigger runaway inference loops, burning GPU cycles and inflating cloud bills without the organization knowing. AIR's pitch is that enterprises need a bill-of-materials for their agent stack, the same way they have one for their software supply chain. If agents become the dominant compute workload by 2028 — and the trajectory suggests they will — this is the security layer the industry cannot skip.

4. The Paper That Made All of This Possible

A Medium post reconstructing the origin story of 'Attention Is All You Need' is circulating widely — and the WTF detail holds up: Aidan Gomez was 20 years old and sleeping in a Google office when he co-authored the paper that became the architectural foundation of every major AI system running today. The hardware significance is underappreciated. The Transformer's self-attention mechanism is nearly perfectly suited to GPU parallelism — each attention head runs as an independent matrix multiply, and modern GPUs have thousands of CUDA cores optimized for exactly that operation. The Transformer didn't just change software; it made GPUs the only viable training hardware for frontier AI. No Transformer paper, no GPU supercycle.

5. Musk's Grok, Colossus, and the Governance Question

Elon Musk directed Grok to produce a targeted roast of Billie Eilish, framing it as political commentary on capitalism and celebrity. The gossip angle is noise. The hardware angle is signal: xAI operates Colossus, a data center housing a massive GPU cluster. Every time Musk points Grok at a political target, that output runs on Colossus infrastructure. The question the AI safety community is now asking explicitly: when the owner of the compute is also the editor of the model, what governance structure exists between the power and the output? The hardware layer is where the power actually sits.

6. China's LLM Giants Diverge on Pricing Strategy

Two major Chinese large-model companies reported interim results revealing opposite strategies: one is raising prices to capture premium enterprise value; the other is cutting aggressively to capture volume. The hardware constraint shapes both bets. Chinese AI labs operate under US export controls restricting access to Nvidia's highest-end chips — H100, H200, and now Blackwell are effectively embargoed. The volume player is betting that Huawei's Ascend 910C and domestically produced alternatives are adequate to serve at scale cheaply. The premium player is betting that whatever restricted Nvidia supply it secured pre-embargo is sufficient to serve high-value customers at margin. This is hardware supply-chain divergence playing out as business strategy in real time — and it's a leading indicator of where global AI pricing settles.

7. Gemini Replaces the Morning Scroll

An Android Police writer replaced their morning social media scroll with Gemini and found it genuinely better. The hardware angle here is the most personal in today's set: Google's Gemini runs on both cloud TPUs for heavy inference and Tensor chips in Pixel devices for on-device tasks. When you ask Gemini something on a modern Pixel device, the phone's on-device NPU handles lightweight inference locally at near-zero latency and zero API cost. When you ask something complex, it routes to the cloud. The UX of swapping your morning scroll is actually an NPU-plus-TPU architecture story — and it explains why Google's vertically integrated silicon strategy gives Gemini a latency and cost edge on the device it ships with.

8. Orchestra Tackles Enterprise Data and AI Management

Orchestra launched a platform aimed at enterprise data and AI pipeline management. The infrastructure angle: modern enterprise AI stacks are a patchwork of compute environments — cloud GPUs for training, edge inference on-prem, orchestration layers connecting heterogeneous hardware. Orchestra's pitch is a unified control plane across that entire surface. From a hardware operations perspective, the real value is visibility into where compute dollars are going across a mixed fleet before the monthly cloud bill arrives as a shock. As enterprises run more parallel AI workloads, the absence of a unified management layer translates directly into runaway compute costs and invisible inefficiencies. Orchestra is positioning itself as the observability layer for that multi-hardware reality.

Quick Hits

• Anthropic's $35B will flow largely into Nvidia Blackwell GPU procurement and data-center power infrastructure — the silicon supply chain just gained a new anchor tenant at historic scale.
• OpenAI's $1B ads run rate is an inference-cost story as much as a revenue story — every personalized response generated for an ad impression burns GPU-seconds.
• China's export-control-constrained LLM labs are stress-testing whether Huawei Ascend can hold up as a Nvidia substitute at volume — the results will reshape the global AI hardware map.
• The Tensor G4 NPU in Pixel 9 is one of the most underrated on-device AI accelerators shipping at consumer scale — Gemini's local-inference latency advantage runs through that chip.

The Anchor

$35 Billion in Silicon: What Anthropic's Raise Actually Buys

Thirty-five billion dollars is an abstraction. Let's make it concrete.

A single Nvidia H200 SXM GPU — among the current top choices for AI training — carries a substantial list price per unit., though hyperscale buyers negotiate below that. A training cluster sufficient to develop a frontier model requires vast numbers of GPUs running continuously for extended periods per training run. At the high end, hardware costs alone reach into the billions — before the data-center buildout for power, cooling, networking, and real estate on top.

So $35 billion doesn't buy a single training run. It buys a multi-year compute roadmap: the clusters to train Claude 5 and Claude 6, the inference infrastructure to serve millions of API calls per day, the redundancy and failover capacity required to meet enterprise SLAs, and the research compute budget to run thousands of ablations and experiments that never reach a final model release.

The supply-chain implications are significant. Anthropic's raise arrives precisely as Nvidia's Blackwell architecture — B100, B200, GB200 — is ramping production at TSMC's most advanced nodes. Every large AI lab placing Blackwell orders is competing for the same TSMC wafer starts, the same CoWoS advanced packaging capacity, and the same HBM3e memory from SK Hynix and Micron. A $35 billion capital commitment signals Anthropic is locking in supply agreements at a scale that reduces available capacity for every other AI lab, cloud provider, and enterprise buyer in the queue. That's not a side effect — it's a strategic consequence.

There's a longer game here as well. Anthropic has been open about its interest in custom silicon — specifically, designing accelerators optimized for Claude's architecture in ways that general-purpose GPUs cannot be. With $35 billion in capital, a custom ASIC program — which requires enormous upfront investment to develop from scratch — becomes genuinely feasible. Google, Amazon, Microsoft, and Meta have each developed their own custom AI silicon. Anthropic has not announced a custom silicon program of its own, unlike several other major frontier labs. That status is unlikely to survive this funding round intact.

The compute arms race has a clear new leader. For now, it runs on Nvidia silicon. But the $35 billion bet suggests Anthropic is already planning for the layer beneath the layer.

Deep Dive

How 'Attention Is All You Need' Made GPUs the Engine of Frontier AI

The Transformer paper — published in 2017 by Vaswani et al., with 20-year-old intern Aidan Gomez as a co-author — is famous for introducing the attention mechanism. What receives far less discussion is exactly why it made GPUs the only viable hardware for frontier AI training. The answer is architectural alignment, and it goes deep.

Before Transformers, the dominant sequence-modeling architectures were recurrent neural networks (RNNs) and LSTMs. These process data sequentially — each timestep depends on the output of the previous one. That sequential dependency is fundamentally incompatible with GPU parallelism. GPUs carry thousands of CUDA cores designed to execute the same operation on many data points simultaneously. But if step N cannot begin until step N-1 completes, you can use only one core at a time. RNNs were GPU-hostile by design, regardless of how much hardware you threw at them.

The Transformer eliminates sequential processing entirely. Self-attention computes relationships between all positions in a sequence simultaneously. In the scaled dot-product attention formula — Attention(Q, K, V) = softmax(QK^T / sqrt(d_k)) * V — the matrix multiplications can be parallelized across every element in the sequence at once. In multi-head attention, each head runs its own independent set of matrix operations. That maps almost perfectly onto GPU architecture: each attention head gets its own CUDA cores; the matrix multiplies run in full parallel; the hardware approaches full utilization.

This alignment is not coincidental — it is structural. Matrix multiplication is the core operation in both self-attention and the feed-forward layers of every Transformer block. It is also the operation that GPU hardware is most aggressively designed to perform at maximum throughput. Nvidia's TensorCores are specifically engineered to accelerate mixed-precision matrix multiply-accumulate operations. The Transformer and the TensorCore arrived simultaneously and fit each other like a key fits a lock.

The result: Transformers scale with compute in a way that RNNs never could. Empirical scaling laws showed that Transformer performance improves predictably as you add more parameters, more training data, and more compute. That gave the entire industry a single, legible roadmap: buy more GPUs, get better models. The GPU supercycle that has reshaped the semiconductor industry, enriched Nvidia's shareholders, and driven Anthropic to raise $35 billion in a single round is a direct downstream consequence of an architectural choice made in a 2017 paper partly authored by a 20-year-old sleeping on an office floor.

For hardware buyers today: every frontier model, every inference API call, every agentic workflow runs on silicon architectures that are direct descendants of GPU optimization patterns unlocked by the Transformer. Nvidia's Blackwell, AMD's MI300X, Intel's Gaudi 3, and Qualcomm's Cloud AI 100 are all, in one sense, hardware answers to a seven-year-old architectural paper. Understanding how attention works is understanding why the silicon industry is shaped the way it is.

One Technique

Estimate Your AI Compute Costs Before You Commit

Before deploying any AI workload — whether a fine-tuned model, an agent pipeline, or a batch inference job — run a back-of-envelope compute estimate first. The formula for inference: (tokens per request) × (requests per day) × (cost per 1K tokens) = daily inference cost. For training or fine-tuning: (model parameters in billions) × (training tokens in billions) × 6 × (cost per FLOP). Most enterprise AI projects overspend significantly because they skip this step and discover costs only after the monthly cloud bill arrives as a shock. Running the numbers upfront lets you select the right model size, the right hardware tier, and the right batching strategy before any capital is committed.

One Prompt

Use this prompt before starting any AI deployment to surface your compute requirements and cost drivers:

I am planning to deploy [describe your AI use case in 2-3 sentences]. Help me estimate the compute requirements before I build.

1. Approximate tokens per request (input + output combined)
2. Expected requests per day at initial launch and at mature scale
3. Recommended model size for this task (in billions of parameters)
4. Estimated daily inference cost at current API pricing for the top two model providers
5. Whether this workload is better served by a cloud API, a hosted open model, or a locally-run model on consumer GPU or NPU hardware
6. One concrete optimization that would reduce compute cost by at least 30% without meaningfully degrading output quality

Assume I am optimizing for cost-efficiency at scale rather than peak raw performance.

One Tip

Match model size to your hardware reality. If your laptop has an NPU — Apple M-series, Qualcomm Snapdragon X Elite, or Intel Core Ultra — a 7B or 8B quantized model running locally will be faster and cheaper than routing to a cloud API for most business tasks. Reserve cloud GPU capacity for workloads that genuinely require 70B+ parameter models. Summarization, classification, first-draft writing, and structured extraction almost never need them. A 7B quantized model on your NPU returns results in under a second at zero marginal cost per query.

Tool of the Day

LM Studio

LM Studio lets you download and run open-source large language models locally on your own hardware — Mac with M-series chips, Windows with Nvidia or AMD GPU, or Linux. It supports quantized models in GGUF format from Hugging Face, provides a ChatGPT-style interface, and exposes a local OpenAI-compatible API endpoint so your existing code connects without changes. Best for: developers and analysts who want to run 7B–13B models at zero marginal cost per query, experiment with open-weight models, or keep sensitive data off cloud APIs entirely. Honest limit: Larger models require a GPU with substantially more VRAM to run effectively.; on most consumer laptops you are effectively limited to quantized 7B–8B models. Available free at lmstudio.ai.

Signature Bites

• The GPU supercycle has a single causal ancestor: a 2017 paper co-authored by a 20-year-old intern who was sleeping in the office the night he helped write it.
• $35 billion in one raise doesn't just buy Anthropic compute — it reserves TSMC wafer capacity and HBM3e memory that every other AI buyer in the queue can no longer access.
• China's LLM pricing divergence is a hardware-constraint story wearing a business-strategy suit: one lab bet on Huawei Ascend, the other held Nvidia supply.
• The NPU in your laptop is almost certainly fast enough for most of your business AI tasks — you are probably paying cloud API costs you do not need to pay.

Joke of the Day

Why did the GPU break up with the CPU?

Because it said: 'You only process one thing at a time. I need someone who can handle thousands of things simultaneously. It's not me — it's your sequential architecture.'

Fact of the Day

A single Nvidia H100 GPU draws substantial power under full AI training load. A GPU cluster at the scale required for frontier model training consumes enormous amounts of power. That is enough electricity to power a significant number of homes simultaneously. The largest AI training clusters now operating consume power at the scale of small cities, and the grid buildout required to serve them is driving a measurable increase in US data-center energy demand projections.

Stat That Matters

$35,000,000,000 — Anthropic's single-round raise, the largest in AI history. For context: Before the AI supercycle peaked, Nvidia's annual revenue was considerably lower than it is today. The scale of Anthropic's recent fundraising reflects just how dramatically the AI capital landscape has shifted. That single comparison tells you more about the magnitude of compute capital forming at the frontier than any press release language can.

Trends

Three macro signals from today's story corpus: Agentic AI is the busiest lane by a wide margin, signaling an industry pivot from model capability to agent deployment — and with it, a shift in compute demand from large batch training runs to millions of small, continuous inference calls requiring different hardware optimization targets. Funding activity is running at a pace suggesting the capital cycle has not peaked; recent agent-security raises are the leading edge of an infrastructure-layer funding wave. China-AI coverage reflects an accelerating domestic hardware race — Huawei's Ascend roadmap, domestic HBM alternatives, and SMIC process improvements — that could produce credible Nvidia alternatives in the coming years if the trajectory holds.

Bold Prediction

Within 18 months, Anthropic announces a custom AI accelerator program — either a full proprietary ASIC or a deep custom silicon co-design with an established chip partner. The $35 billion raise is the capital event that makes it feasible. The precedent is unambiguous: Google built TPUs, Amazon built Trainium, Microsoft built Maia, Meta built MTIA. Anthropic is currently the only major frontier lab without a silicon program. That status does not survive this funding round intact. When the announcement comes, it will be framed as a research efficiency story — but the real motivation is breaking long-term dependence on Nvidia's supply constraints and pricing power.

Paper Watch

FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness.

Standard Transformer attention requires memory that scales quadratically with sequence length. For long documents, that blows out GPU VRAM and forces the model to break the sequence into chunks, losing the ability to attend across the full context. FlashAttention restructures the computation using a technique called tiling: it blocks the attention matrix into chunks that fit in GPU SRAM — the fast on-chip cache — rather than repeatedly reading from and writing to slower HBM (high-bandwidth memory). The result is significantly faster attention computation and a substantially reduced memory footprint., and the ability to support much longer context windows without requiring larger GPUs. Every major model serving long contexts today runs FlashAttention or a direct derivative. It is the reason you can paste a 100-page document into Claude and receive a response in seconds rather than minutes. Understanding FlashAttention means understanding how AI systems actually run on the hardware layer beneath every API call.

Founder Spotlight

The AIR Team — Naming a New Category

The founders of AIR just closed $50 million to build agent supply-chain security — vetting the skills, plug-ins, and runtime add-ons that enterprise AI agents consume. The strategic move here is category creation, not product launch. AIR isn't simply selling a tool; it's naming a problem that most enterprises have not yet formally articulated: how do we audit what our agents are actually running at inference time, and who approved it? In a market where agent deployment is accelerating faster than governance frameworks can follow, whoever defines the category vocabulary tends to own the category. The $50 million signals that serious enterprise buyers are already treating agent skill vetting as a compliance and security requirement rather than a nice-to-have feature. Watch for a cluster of similar raises in the agent infrastructure security space over the next 12 months — AIR has just marked the starting line.

Quote

'One demands higher prices upward — the other seeks greater volume downward.'

— 21财经, on the diverging strategies of China's two leading large-model companies, as constrained by hardware access and export controls.

Learner's Edge

Training Compute vs. Inference Compute

AI systems consume hardware in two fundamentally different modes, and confusing them leads to poor infrastructure decisions.

Training compute is the one-time cost of teaching a model: feeding it billions of examples, adjusting billions of parameters, running for weeks or months on clusters of thousands of GPUs. It is expensive, highly parallel, and complete before any user ever interacts with the model. Training favors giant clusters, high-bandwidth GPU-to-GPU interconnects like NVLink and InfiniBand, and maximum FLOP throughput.

Inference compute is the ongoing cost of using a trained model: every query you type generates a GPU response. Per-operation it is cheaper than training — but it happens millions of times per day across all users, and at scale inference costs routinely exceed training costs for deployed models. Inference favors low-latency, cost-efficient hardware closer to users: on-device NPUs, smaller GPU instances, and batching strategies that amortize overhead.

Anthropic's $35 billion covers both modes — and the allocation between them reveals a great deal about the company's growth assumptions for both research velocity and commercial deployment scale.

Sign-off

That's The Foundry for today. The chips are moving — and now you know why. See you tomorrow.

Sources

1. Anthropic just made a staggering $35 billion bet on Claude — here's why it needs so much power
2. OpenAI’s ChatGPT Ads Business Hits $1bn Annualized Revenue Run Rate
3. AIR raises $50M to help companies vet the skills and add-ons AI agents use
4. Attention Is All You Need: The Translation Problem That Led to ChatGPT — abdullahsaad5.medium.com
5. Elon Musk asks Grok to roast Billie Eilish over capitalism and luxury
6. Divergence in Interim Reports of Large Model Rivals: One Demands Higher Prices Upward, the Other Seeks Greater Volume Downward
7. Swapping my morning scroll for Gemini turned out to be a brilliant move
8. Orchestra launches management solution for enterprise data and AI