Google has done something it avoided for months: it put a real, published dollar figure on its newest piece of hardware. On September 8, 2026, Google Cloud’s pricing documentation shows the seventh-generation Tensor Processing Unit, code-named Ironwood, listed at $12.00 per chip-hour in Iowa and $13.20 per chip-hour in London. For a chip that reached general availability back in April and then sat without a public rate for three months, that’s the detail buyers have been waiting for.
The timing matters. Nvidia still dominates the AI accelerator market, but its biggest customers are being told to expect higher costs on next-generation Grace Blackwell and Vera Rubin systems as memory prices climb. Google’s decision to make Ironwood’s economics visible, tier by tier, region by region, turns a strategic guessing game into a side-by-side spreadsheet exercise for every cloud AI buyer running procurement numbers this quarter.
Google Finally Puts a Price Tag on Ironwood TPU
Ironwood is Google’s seventh-generation TPU, also referred to in some cloud documentation as TPU7x. According to TECHi, the chip reached general availability on April 22, 2026, yet Google published no on-demand list price for roughly three months afterward, while older generations like Trillium (v6e) kept their public rates intact the entire time.
That gap is now closed. Google’s own Cloud TPU pricing page lists Ironwood at $12.00 per chip-hour on demand in the us-central1 (Iowa) region and $13.20 per chip-hour in europe-west2 (London). Flexible capacity through Dynamic Workload Scheduler Flex-start pricing comes in lower, at $6.00 per hour in both regions, and buyers willing to commit for one or three years get steeper discounts still: $8.40 and $5.40 per hour in Iowa on one-year and three-year terms, respectively, and $9.24 and $5.94 per hour in London.
The rate is delivered at the TPU VM level rather than as loose single-chip retail inventory. Google sells Ironwood in fixed VM shapes such as tpu7x-standard-4t, a four-chip configuration, according to Google Cloud’s own release notes, which keeps pricing and provisioning tied to cluster-friendly units instead of arbitrary chip counts.
Why Three Months of Silence Actually Worked in Google’s Favor
Keeping a chip generally available but unpriced isn’t an accident for a company that runs one of the three biggest public clouds. TECHi’s reporting frames the silence as a negotiating tool: without a published rate, Google could cut individual deals with the customers it wanted most, rather than exposing a floor price that every prospective buyer would immediately try to beat down.
That’s reportedly what happened with Anthropic. A SemiAnalysis estimate cited by TECHi put Anthropic’s negotiated rate for Ironwood capacity at roughly $1.60 per TPU-hour, a fraction of the now-published $12.00 on-demand price. That’s not a public list price; it’s a large, strategically important customer’s contract rate. But it tells a story about the spread between what Google will charge a marquee AI lab chasing training capacity and what everyone else pays without leverage.
Publishing the on-demand and committed-use rates now doesn’t eliminate that gap. It just makes the top of the range visible, which is exactly what mid-size AI companies and enterprise buyers have been asking cloud reps for since Ironwood launched.
Ironwood Pricing Table: On-Demand, Flex-Start and Committed Rates
Here’s the full published rate card, per chip-hour, as it appears on Google Cloud’s TPU pricing page:
| Pricing Tier | Iowa (us-central1) | London (europe-west2) |
|---|---|---|
| On-demand | $12.00/hr | $13.20/hr |
| DWS Flex-start | $6.00/hr | $6.00/hr |
| 1-year commitment | $8.40/hr | $9.24/hr |
| 3-year commitment | $5.40/hr | $5.94/hr |
| Reported negotiated rate (Anthropic, per SemiAnalysis) | ~$1.60/hr (contract-specific, not public list price) | |
The spread between the three-year committed rate and the on-demand rate is more than 2x in both regions. That’s a familiar cloud pricing pattern, but it’s a meaningful one for teams deciding whether to lock in Ironwood capacity now or wait for the next chip cycle.
Inside Ironwood: Memory, Bandwidth and Superpod Scale
The pricing news lands alongside a clearer picture of what Ironwood actually is under the hood. According to reporting from Chosun Biz, each Ironwood chip carries 192GB of HBM per chip, roughly matching the memory capacity Nvidia ships on its own newest data center GPUs. Separate technical comparisons put Ironwood’s memory bandwidth at approximately 7.37 TB/s per chip.
Where Ironwood differentiates itself is at cluster scale. Google’s design reportedly lets Ironwood pods scale up to 9,216 chips in a single superpod, with roughly 1.77 petabytes of shared HBM and 9.6 Tb/s of inter-chip interconnect bandwidth. That’s a pod-scale inference architecture, not a general-purpose rentable GPU, a distinction that shows up directly in how Google restricts access. Ironwood is sold exclusively through Google Cloud. There’s no third-party marketplace listing, no bare-metal resale, and no multi-cloud availability the way Nvidia’s H100 and B200 show up across dozens of GPU cloud providers.
How Ironwood Stacks Up Against Nvidia’s H100 and B200
Direct, apples-to-apples throughput benchmarks between Ironwood and Nvidia’s Blackwell-generation chips aren’t cleanly published yet, so any comparison has to lean on what is public: memory, bandwidth, and rental economics. On that front, the picture is clearer than the marketing usually allows.
Nvidia’s H100 SXM, still the most widely deployed accelerator in production AI infrastructure, ships with 80GB of HBM3 and roughly 3.35 TB/s of memory bandwidth per Nvidia’s own published specifications. On the GPU marketplace Spheron, H100-class capacity has traded in the low single digits per hour depending on provider and commitment. Nvidia’s newer B200, built on the Blackwell architecture, carries 192GB of HBM3e (matching Ironwood’s per-chip capacity) and roughly 8 TB/s of bandwidth. Spheron’s own live marketplace data, checked as of July 7, 2026, listed B200 SXM6 on-demand pricing at $9.36 per GPU-hour, with spot pricing running $5.34 to $5.37 per GPU-hour.
Set next to that, Ironwood’s $12.00 on-demand rate in Iowa looks higher than B200’s on-demand list price on at least one third-party marketplace, though Ironwood’s committed-use rates undercut B200 once a buyer locks in a one- or three-year term. The honest takeaway, echoed by the analysis at ValueAddVC, is that Ironwood isn’t positioned as a cheaper GPU alternative. It’s a Google-Cloud-only, pod-scale inference chip aimed at customers already committed to Google’s stack who want predictable, high-memory capacity at a scale Nvidia’s rental market doesn’t offer in the same packaging.
Competitive Snapshot: TPU v7 vs H100 vs B200 vs MI300X
| Chip | Memory | Bandwidth | On-Demand Price | Availability |
|---|---|---|---|---|
| Google TPU v7 Ironwood | 192GB HBM3e | ~7.37 TB/s | $12.00/hr (Iowa) | Google Cloud only |
| Nvidia H100 SXM | 80GB HBM3 | ~3.35 TB/s | ~$2.69-3.00/hr (aggregator avg.) | Multi-cloud, widely resold |
| Nvidia B200 SXM6 | 192GB HBM3e | ~8 TB/s | $9.36/hr on-demand; $5.34-5.37/hr spot | Multi-cloud marketplaces |
| AMD MI300X | 192GB HBM3 | ~5.3 TB/s | Varies by provider, not uniformly published | Microsoft, Meta and select clouds |
Memory capacity is where Google, Nvidia’s B200 and AMD’s MI300X have effectively converged on 192GB per accelerator. The differentiators now sit in bandwidth, interconnect design, and, increasingly, who will actually sell a buyer the chip and on what terms.
The AI Chip Market Share Fight: Nvidia, AMD and Google
None of this pricing news changes the market structure overnight. Nvidia’s dominance remains enormous. ValueAddVC’s 2026 market analysis puts Nvidia’s share of the AI accelerator market at roughly 73% to 80%, alongside a reported $89.0 billion in quarterly data center revenue. A separate estimate from GoMarkets, citing IDC, puts Nvidia’s control of the AI data center chip market at roughly 81%, a lead built in part on record quarters like the $96.2 billion result Nvidia posted even as rivals circle.
Google TPU sits well behind, at an estimated 6% to 8% of the broader AI chip market measured by deployed FLOPS, concentrated almost entirely inside Google Cloud and Google’s own internal workloads rather than sold as general-purpose hardware. AMD has climbed to roughly 5% to 7% share, driven by MI300X and MI325X inference deployments, with Microsoft and Meta named as its largest customers, while custom silicon efforts like Microsoft’s Maia 300 add another layer of pressure on Nvidia from the hyperscaler side.
Google’s TPU camp isn’t standing still on the customer side either. ValueAddVC’s reporting names Anthropic and Meta as active TPU users running workloads on both TPU v6 Trillium and the newer v7 Ironwood, a signal that Google is winning specific, high-volume inference and training contracts even while its overall share of the market stays in the single digits.
What makes Ironwood’s pricing news notable isn’t that it will flip those market share numbers in a single quarter. It’s that Google is now competing on a dimension, published, predictable, multi-year cloud pricing, where Nvidia has historically had the advantage simply by virtue of selling hardware outright to dozens of resellers who then compete on price against each other. A single vendor publishing a locked three-year rate is a different kind of competitive pressure than a fragmented marketplace undercutting itself.
Nvidia’s Own Cost Pressure Adds Context
Google’s pricing move doesn’t happen in a vacuum. Nvidia’s biggest customers have reportedly been warned that AI server builds around Grace Blackwell and the upcoming Vera Rubin platform could cost more than 15% extra starting in early 2027, driven by rising memory prices across the industry, a warning that surfaced in Yahoo Finance’s markets coverage in late August 2026 and echoes the AI server price hikes Nvidia has already pushed through on memory grounds. That’s the same memory crunch that’s been pushing DRAM and HBM contract prices sharply higher industry-wide throughout 2026.
The irony is that Google isn’t immune to the same pressure. Ironwood’s 192GB of HBM3e per chip means Google is buying from the exact same constrained memory supply chain as Nvidia and AMD. What’s different is who absorbs the volatility. By publishing fixed one- and three-year rates now, Google is effectively betting that it can manage its own memory procurement well enough to honor those locked prices even if spot HBM costs move against it. Nvidia’s GPU cloud partners, by contrast, tend to pass rising component costs through to renters faster, since most of that capacity trades on shorter-term, spot-influenced marketplace pricing rather than a single vendor’s multi-year rate card.
For buyers comparing Ironwood’s newly published, fixed-through-commitment pricing against Nvidia hardware whose future cost trajectory is less certain, the calculus shifts. A three-year Ironwood commitment at $5.40 per hour in Iowa locks in a rate today that doesn’t move if HBM prices climb again next year. Nvidia GPU rental pricing, by contrast, is set by a competitive multi-cloud marketplace that can and does reprice as component costs change.
From TPU v1 to Ironwood: A Decade of Custom Silicon
Google has now shipped seven generations of its own AI accelerator, a program that started as an internal-only project to speed up inference for services like Search and Translate before Google began renting TPU capacity to outside Cloud customers. Trillium (TPU v6e) was the workhorse generation that carried most of Google’s external TPU business through 2025, and it’s notable that Google kept Trillium’s public pricing intact even while Ironwood launched without one, according to TECHi’s reporting.
That generational overlap, an older chip with transparent pricing running alongside a newer one priced opaquely for months, is a pattern Google has used before to manage capacity allocation between existing customers and the labs racing for the newest hardware. Ironwood’s move to full published pricing marks the point where Google is treating the seventh generation as ready for broader, self-service adoption rather than reserved primarily for negotiated enterprise deals.
Market Impact: What This Means for Cloud AI Buyers
For engineering teams sizing out training or inference budgets this quarter, the practical impact is straightforward: Ironwood’s economics can now be modeled without a Google sales call. A team that would have had to negotiate blind now has published on-demand, flex, one-year and three-year numbers to run against Nvidia and AMD marketplace rates.
That said, the single-cloud restriction remains the biggest practical constraint. Any team not already standardized on Google Cloud has to weigh Ironwood’s per-chip economics against the cost and friction of migrating workloads and data. For teams already inside Google’s ecosystem, particularly Anthropic and Meta-scale inference operations, the newly public committed-use pricing gives finance and infrastructure teams a real number to plan against instead of an estimate pulled from a leaked contract.
Checking Current TPU Pricing and Availability
Teams evaluating Ironwood capacity can pull current TPU VM availability directly through the Google Cloud CLI before committing to a pricing tier:
# List available TPU accelerator types in a given zone
gcloud compute tpus accelerator-types list --zone=us-central1-a
# Check current TPU VM instances and their configured type
gcloud compute tpus tpu-vm list --zone=us-central1-a
# View the reservation and commitment options tied to your billing account
gcloud compute commitments list --region=us-central1Running these before signing a one- or three-year commitment is worth the extra step, since regional availability for tpu7x-standard shapes has varied by zone since launch.
Predictions: Where the TPU-Nvidia Price War Goes Next
A few things look likely to play out over the next two to three quarters. First, expect Nvidia’s cloud partners to respond to Ironwood’s published committed-use rates with more aggressive multi-year discounting of their own, especially on B200 and Blackwell-generation capacity, to keep price-sensitive customers from evaluating a Google Cloud migration. Second, Google will likely keep negotiating below-list rates for anchor customers like Anthropic even with a public price sheet now in place, since the $1.60-per-hour reported rate shows how much room Google has to discount for volume without touching its published numbers.
Third, the memory shortage driving Nvidia’s reported 15%-plus cost increases into 2027 isn’t going away quickly, which means the gap between Ironwood’s fixed three-year rate and Nvidia’s marketplace-driven pricing could widen further in Google’s favor if HBM and DRAM costs keep climbing. Fourth, AMD’s MI300X and MI325X will keep picking up inference-specific deals at Microsoft and Meta rather than challenging Nvidia head-on for frontier training workloads, keeping AMD’s share in the mid-single digits rather than triggering a major share shift, even as more of Nvidia’s own biggest customers keep building rival chips in-house. Fifth, don’t expect Google to open Ironwood to third-party resale or multi-cloud availability in the near term. The pod-scale architecture and Google-Cloud-only packaging suggest Google is optimizing for retention of its own customers, not for competing broadly in the GPU rental marketplace Nvidia and AMD chips already dominate.
What to Watch Next
The next real signal will come from Google’s next earnings call and from whether Anthropic, Meta or other large TPU customers publicly disclose changes to their Ironwood usage now that list pricing exists. Watch also for whether Google extends published pricing to additional regions beyond Iowa and London, and whether Nvidia or its cloud resellers respond with matching multi-year commitment discounts on B200 capacity, especially after H100 and B200 cloud rental rates already moved sharply this year. Any move by Google to sell Ironwood capacity through a third party, however unlikely in the short term, would be the biggest structural shift this story could produce.
Frequently Asked Questions
What is Google’s Ironwood TPU?
Ironwood is Google’s seventh-generation Tensor Processing Unit, also referred to as TPU7x in some documentation, designed for large-scale AI training and inference workloads on Google Cloud.
How much does the Ironwood TPU cost per hour?
Google Cloud’s published on-demand rate is $12.00 per chip-hour in the Iowa (us-central1) region and $13.20 per chip-hour in London (europe-west2), with lower rates available through Flex-start pricing and one- or three-year commitments.
Can I rent Ironwood TPUs outside of Google Cloud?
No. Ironwood is currently sold exclusively through Google Cloud, unlike Nvidia’s H100 and B200, which are available across numerous multi-cloud GPU marketplaces.
Is Ironwood cheaper than Nvidia’s B200?
Not on a pure on-demand basis in every market. One GPU marketplace, Spheron, listed B200 on-demand pricing at $9.36 per hour as of July 2026, below Ironwood’s $12.00 Iowa rate. Ironwood’s committed-use discounts, down to $5.40 per hour on a three-year term, can undercut B200’s on-demand rate for buyers willing to commit long term.
Why did Google wait months to publish Ironwood pricing?
According to TECHi’s reporting, the chip reached general availability in April 2026 but went without a public on-demand price for roughly three months, a gap that let Google negotiate individual rates with large customers, reportedly including a rate near $1.60 per hour for Anthropic, before exposing a public floor price to the wider market.
How much AI chip market share does Google TPU actually have?
Estimates put Google TPU at roughly 6% to 8% of the broader AI chip market by deployed FLOPS, compared with an estimated 73% to 81% for Nvidia and 5% to 7% for AMD, according to 2026 market research cited by ValueAddVC and GoMarkets.
Which companies are using Google’s TPU v7 Ironwood?
Anthropic and Meta have been named as active users of Google’s TPU v6 Trillium and v7 Ironwood chips, according to ValueAddVC’s 2026 AI hardware market analysis.
Does the memory shortage affect TPU pricing too?
Ironwood’s per-chip 192GB of HBM3e means Google is exposed to the same memory market as Nvidia and AMD. While Google’s committed-use rates are now fixed for one or three years, Nvidia’s customers have reportedly been warned of cost increases of more than 15% on Blackwell and Vera Rubin systems into 2027 as memory prices rise.
