Google TPU v7 Ironwood vs NVIDIA B200: Inference Cost (2026)

Google's TPU v7 Ironwood reached general availability on April 22, 2026. More than two months later, Google still hasn't put a price on its own pricing page. That's unusual for a chip Google has been selling to some of its biggest customers since spring, and it leaves anyone comparing TPU v7 Ironwood vs NVIDIA B200 doing the math with half the inputs missing.

This post works through what's actually confirmed about Ironwood, the one real dollar figure that has surfaced anywhere, and an honest cost-per-token comparison against B200, built from live pricing rather than a guess. Where a number can't be verified, it's flagged as an estimate, not presented as fact.

What Makes Ironwood (TPU v7) Different From Trillium and TPU 8i

Ironwood is not an upgrade to Trillium (TPU v6e), and it's not the same chip as TPU 8i. All three show up in searches for "Google TPU," and it's easy to conflate them if you're not tracking Google's chip numbering closely. Here's where each one sits.

Specs at a Glance: HBM, FP8, and Interconnect

Ironwood is the first TPU generation with native FP8 hardware support. Trillium topped out at BF16. That single change is why Ironwood's headline FP8 number looks so much bigger than anything Google published for the prior generation, and why comparing it to B200's FP8 throughput is a fair fight rather than an apples-to-oranges one.

A full Ironwood pod scales to 9,216 chips for a combined 42.5 FP8 ExaFLOPS. Google says that's a 10x peak performance jump over TPU v5p and more than 4x the performance per chip of Trillium, for both training and inference. Those are Google's own figures; no third party has independently reproduced them on a named model yet.

Why This Is a Different Generation Than the Site's Two Existing TPU Comparisons

We've covered Trillium v6 against B200 and TPU 8i against Rubin and B200 on this blog. Ironwood sits between them in Google's roadmap, and it's the generation actually running production traffic right now, which is exactly why it deserves its own numbers instead of being folded into either of those posts.

Trillium (v6e) is the prior generation: BF16-only, available since 2025, and the chip most of Google's published TPU pricing still refers to. TPU 8i is the next generation after Ironwood, an inference-specialized v8 chip Google previewed at the same Cloud Next 2026 event where Ironwood shipped, targeting a late-2027 GA. Ironwood (v7) is the one chip of the three that is both generally available today and has a real, if incomplete, cost data point attached to it. For the wider hyperscaler-silicon picture beyond Google, see Trainium 3, TPU Ironwood, Maia 200, and MTIA compared.

Ironwood Pricing and Availability as of Mid-2026

Ironwood shipped to production faster than it shipped a public price. That gap is the whole story of buying it right now.

General Availability Timeline and Who Is Buying It

Google announced Ironwood's general availability at Google Cloud Next 2026 on April 22, 2026, alongside a preview of the next-generation split into TPU 8t (training) and TPU 8i (inference). Anthropic is the confirmed anchor customer, and the scale involved is unusual even by hyperscaler standards: an initial phase of 400,000 units (roughly $10B in finished racks purchased through Broadcom), plus a further commitment of around 600,000 more units, tracked as roughly $42B in remaining performance obligations, rented through Google Cloud. Combined, that's up to 1 million TPU chips and more than a gigawatt of capacity committed for 2026.

James Bradbury, Anthropic's Head of Compute, described the deal this way: "Ironwood's improvements in both inference performance and training scalability will help us scale efficiently while maintaining the speed and reliability our customers expect."

Why Google Still Has Not Published a Public Chip-Hour Rate

Google's own TPU pricing page still lists on-demand and committed-use rates for v6e and v5p. Ironwood doesn't appear there. That's consistent with how Google has rolled out new TPU generations before: capacity goes to committed, negotiated deals first, and self-serve list pricing follows later, once supply catches up with the largest customers. For Ironwood, the largest customer took most of the first generation of capacity, which likely explains why there's still nothing for smaller GCP accounts to click "buy" on.

Whatever the reason, the practical effect is the same for anyone not negotiating at Anthropic's scale: there is no rate card to check, no self-serve quote, and no way to run Ironwood today without Google's direct involvement.

The One Real Pricing Data Point That Exists: Anthropic's Negotiated GCP Rate

SemiAnalysis estimates Anthropic is paying roughly $1.60 per TPU-hour on GCP for Ironwood capacity. That number comes from analysis of the deal's disclosed dollar totals and unit counts, not from a Google price sheet, and it reflects a single customer's negotiated large-scale rate, not a list price any other buyer could expect to match.

It's still useful as an anchor. The only publicly listed TPU prices Google has ever put out are for older chips: Trillium (v6e) on-demand runs about $2.70/chip-hour, dropping to $1.89 with a 1-year commitment and $1.22 with a 3-year commitment, and the older v5p starts around $4.20/chip-hour on-demand. If Ironwood eventually gets a public rate, history suggests it'll land somewhere between Anthropic's negotiated $1.60 and a higher on-demand sticker price, the same way Trillium's on-demand rate sits well above its multi-year committed rate.

Cost Per Token: Ironwood vs NVIDIA B200 for LLM Inference

This is the part that actually matters for a buying decision, and it's also where the gap between "confirmed" and "estimated" gets sharpest.

Live B200 Pricing on Spheron Right Now

B200's numbers are not in question. As of July 7, 2026, NVIDIA B200 SXM6 on Spheron runs $9.36/GPU/hr on-demand and roughly $5.34-5.37/GPU/hr spot, pulled directly from live marketplace data, split by instance type before taking a minimum so the on-demand figure isn't accidentally a spot rate in disguise.

Running Llama 70B at FP8 with vLLM, a B200 delivers approximately 1,000 tokens/sec per GPU at batch 8, based on MLPerf Inference v6 data and public vLLM team benchmarks. Turning that into cost per million tokens (CPM):

Pricing fluctuates based on GPU availability. The prices above are based on 07 Jul 2026 and may have changed. Check current GPU pricing → for live rates.

Building an Honest Cost-Per-Token Estimate for Ironwood (and Why You Can't Yet)

A real Ironwood CPM number needs two inputs: a price per chip-hour and a measured tokens/sec figure on a comparable model. We have neither cleanly. The $1.60/hour figure is a single negotiated rate at a scale nobody else operates at, and there is no independent, third-party Ironwood inference throughput benchmark published anywhere as of mid-2026. Any tokens/sec number for Ironwood you see quoted elsewhere is either Google's own internal claim or someone's extrapolation from FLOPS, neither of which substitutes for a measured result on a named model and batch size.

What we can do honestly is work the math backward. Using Anthropic's reported $1.60/hour against B200's CPM figures above, here's the tokens/sec Ironwood would need per chip just to match B200, not beat it:

• To match B200 on-demand ($2.60/M tokens): Ironwood needs roughly 171 tokens/sec per chip
• To match B200 spot ($1.48/M tokens): Ironwood needs roughly 300 tokens/sec per chip

Given Google's claim of more than 4x the performance per chip of Trillium, clearing 171-300 tokens/sec on a chip purpose-built for inference is not a high bar on paper. But "not a high bar on paper" and "confirmed by a benchmark" are different claims, and only the first one is available right now. Treat this as the threshold Ironwood has to clear, not proof that it does.

TCO Claims vs Retail Price: What Google's 44% Number Does and Doesn't Tell a Buyer

Google's own accounting puts Ironwood's all-in total cost of ownership, in a full 3D Torus pod configuration, at roughly 44% lower than the TCO of an NVIDIA GB200 server. That figure is real, but it answers a different question than "what will I pay per hour." TCO folds in power, cooling, networking, and rack-level efficiency across an entire pod that Google itself owns and operates at gigawatt scale. It's an internal engineering comparison, not a retail price signal, and it says nothing about what a chip-hour costs a customer renting capacity rather than owning the infrastructure.

The gap between a 44% TCO advantage and an actual dollar figure is exactly the gap this post keeps running into: Google can show you why Ironwood is efficient to operate at Google's own scale, without showing you what it costs to rent a slice of that efficiency.

FP8-Only vs FP4: The Software-Stack Gap That Changes the Real Number

Ironwood's FP8 support is new for TPUs, but it's still one precision level behind where NVIDIA's Blackwell generation already operates. B200 supports FP4 through the Blackwell Transformer Engine, at up to 9,000 dense FP4 TFLOPS against 4,500 dense FP8 TFLOPS, and TensorRT-LLM and vLLM have shipped FP4 quantization support since early 2026. A B200 workload running FP4-quantized inference at scale can beat the FP8 numbers used above, which means the CPM gap Ironwood needs to close is potentially larger than the FP8-vs-FP8 comparison suggests, depending on whether your model tolerates FP4 quantization accuracy loss.

The other half of the gap is ecosystem, not silicon. Ironwood runs on JAX, MaxText, and Jetstream. None of the standard CUDA inference stack, vLLM, SGLang, TensorRT-LLM, runs on TPU hardware. Porting a production vLLM deployment to JAX is a real engineering cost that doesn't show up in any chip-hour comparison, and it's the same migration tax covered in the Trillium and TPU 8i posts linked above.

When to Wait for Ironwood vs Rent B200 Today

For most teams, this isn't a close call right now, and the reason has nothing to do with which chip is theoretically faster.

Decision Checklist

• You need inference capacity this quarter: rent B200. It's live, self-serve, and priced. Ironwood has no public rate card to check.
• You're already committed to Google Cloud and JAX/MaxText: worth calling your Google account team to ask about Ironwood pricing directly. You won't find it published, but you may be able to negotiate a quote.
• Your workload needs FP4 quantization: stay on GPU cloud. Ironwood is FP8-only; B200 with vLLM or TensorRT-LLM FP4 support is the more mature path today.
• Your commitment is nowhere near Anthropic's scale: don't expect Anthropic's $1.60/hour rate. That number reflects a gigawatt-scale, multi-year deal, not a starting quote for a smaller GCP account.
• You run vLLM, SGLang, or any CUDA-based serving stack: the migration cost to JAX is real and rarely pencils out unless you're already operating at hyperscale. See the migration checklist in the Trillium comparison for what actually has to be rewritten.
• You want to benchmark before committing either way: you can do that on B200 today. You can't yet on Ironwood, since there is no self-serve access and no independent benchmark to check your own numbers against.

For teams working through the general cost-per-token math across GPU options, the cost-per-token benchmark methodology and broader inference cost economics posts cover the same CPM formula used above applied to a wider set of models and hardware. If your workload runs Mixture-of-Experts models, the systolic-array-vs-GPU tradeoffs matter even more than they do for dense models, covered in MoE inference optimization on GPU cloud. And if you're weighing Blackwell-generation GPUs specifically, B300 vs B200 runs the same cost-per-token math one tier up. For deployment specifics once you've picked a GPU, the Spheron docs cover getting a vLLM stack running end to end.

Ironwood's specs are real and its FP8 throughput is a genuine step up for TPUs, but a spec sheet without a price tag isn't something most teams can actually buy against. B200 on Spheron ships with a published rate, native vLLM support, and FP4 quantization today.

B200 on Spheron → | Check live GPU pricing →