Why is it too early to see hyperscaler chips as broadly available GPU replacements?
The evidence shows real progress rather than merely announced silicon: Google trained and served Gemini on TPUs, AWS put nearly 500,000 Trainium2 chips into live Project Rainier, and Meta reported hundreds of thousands of deployed MTIAs. Commitments have also reached multiple gigawatts. But those facts do not establish broad access across customers or on-time execution: Microsoft delayed its next Maia chip after design setbacks and described outside Maia 200 customers as a future prospect. Crux AI launched as a Google-Blackstone TPU cloud venture, but aimed to rent Google processors to customers in 2027, while OpenAI said Jalapeño capacity would serve its own compute demand for a long time. Large deployments and supplier roadmaps therefore show momentum, not yet a broadly available substitute for merchant GPUs.
Trace the chain
Evidence of scale
- 1TPUs run GeminiLaunch
Google's TPU v5p arrives with Gemini, which was trained and served on TPUs
The same day it announced Gemini, Google Cloud introduced TPU v5p, its most powerful accelerator to date, with 8,960 chips per pod, more than twice the FLOPS and three times the HBM of TPU v4, and said Gemini had been trained on and was served from TPUs. It set the pattern for the next three years: each Google model generation ships on a new TPU.
Google Cloud Blog ·
Today, to address these challenges, we are excited to announce Cloud TPU v5p, our most powerful, scalable, and flexible AI accelerator thus far.
Why it matters here. Gemini's TPU training and serving established production use of a hyperscaler accelerator, which Ironwood extended with a TPU designed specifically for inference.
Custom AI Silicon at the Hyperscalers: TPU, Trainium, Maia, MTIA and the OpenAI–Broadcom Chip
- 2Ironwood targets inferenceLaunch
Ironwood: the seventh TPU, built for inference, scales to 9,216 chips and 42.5 exaflops
At Cloud Next Google introduced Ironwood as the first TPU designed specifically for inference and 'thinking' models, with 192 GB of HBM per chip and pods of 9,216 chips delivering 42.5 exaflops. It became the chip behind Google's TPU sales push to Anthropic and Meta over the following year.
Google (The Keyword) ·
Today at Google Cloud Next 25, we’re introducing Ironwood, our seventh-generation Tensor Processing Unit (TPU) — our most performant and scalable custom AI accelerator to date, and the first designed specifically for…
Why it matters here. Ironwood was Google's first TPU designed specifically for inference and supported 9,216-chip pods; the next entry shows AWS operating another custom accelerator, Trainium2, at large scale for Claude training and serving.
Custom AI Silicon at the Hyperscalers: TPU, Trainium, Maia, MTIA and the OpenAI–Broadcom Chip
- 3Rainier operates at scaleLaunch
Project Rainier is live: nearly 500,000 Trainium2 chips, Claude on more than a million by year-end
AWS said Rainier was fully operational less than a year after it was announced, spread across multiple U.S. data centers, with Anthropic already training and serving Claude on it and expected to be on more than one million Trainium2 chips by the end of 2025. It was the largest deployment of a non-Nvidia accelerator for a frontier model.
Amazon ·
Project Rainier, one of the world's largest AI compute clusters, is now fully operational, less than one year after it was first announced.
Why it matters here. Rainier's large live Trainium deployment showed that custom-accelerator scale was operational, while Google's next entry still described Ironwood general availability as due only within weeks.
Custom AI Silicon at the Hyperscalers: TPU, Trainium, Maia, MTIA and the OpenAI–Broadcom Chip
- 4Ironwood availability plannedLaunch
Ironwood reaches general availability, with Claude named alongside Gemini as a TPU-trained model
Google Cloud said Ironwood would be generally available within weeks, with 10x the peak performance of TPU v5p and more than 4x Trillium per chip, and introduced Arm-based Axion N4A instances. The post opened by listing Anthropic's Claude with Gemini, Veo and Imagen as models that train and serve on TPUs.
Google Cloud Blog ·
Ironwood, our seventh generation TPU, will be generally available in the coming weeks.
Why it matters here. Google said Ironwood would be generally available within weeks, while Meta next reported hundreds of thousands of MTIA chips deployed in production.
Custom AI Silicon at the Hyperscalers: TPU, Trainium, Maia, MTIA and the OpenAI–Broadcom Chip
- 5Meta reports installed MTIAsAnnouncement
Meta lays out four MTIA generations in two years: 300, 400, 450 and 500, on a six-month cadence
Meta said it had deployed hundreds of thousands of MTIA chips, renamed the first two parts MTIA 100 and 200, and published a roadmap of four more generations through 2027 that widens the chips from recommendation inference to recommendation training and generative-AI inference, built with Broadcom on modular chiplets. The post committed Meta to iterating silicon at a consumer-hardware pace.
Meta AI Blog ·
Since introducing MTIA 100 and 200, we have accelerated MTIA development across four successive generations: MTIA 300, 400, 450, and 500. These new chips have either already been deployed or are scheduled for…
Why it matters here. Meta reported hundreds of thousands of MTIA chips deployed in production, while Anthropic's next agreement committed up to 5 GW of Trainium capacity.
Custom AI Silicon at the Hyperscalers: TPU, Trainium, Maia, MTIA and the OpenAI–Broadcom Chip
- 6Trainium reaches five gigawattsAnnouncement
Amazon puts up to $25 billion more into Anthropic; Anthropic commits $100 billion to AWS and up to 5 GW of Trainium
Amazon invested $5 billion immediately with up to $20 billion more tied to milestones, on top of $8 billion since 2023, and Anthropic committed more than $100 billion over ten years to AWS, securing up to 5 GW across Trainium2, 3 and 4 with nearly 1 GW of Trainium2 and Trainium3 online by end-2026. Anthropic said it already used more than one million Trainium2 chips and kept AWS as its primary training provider.
Anthropic ·
We have signed a new agreement with Amazon that will deepen our existing partnership and secure up to 5 gigawatts (GW) of capacity for training and deploying Claude, including new Trainium2 capacity coming online in…
Why it matters here. Anthropic committed up to 5 GW of Trainium capacity, including capacity planned through the end of 2026; Broadcom next reported commercial XPU revenue while laying out additional TPU capacity scheduled for 2027 and 2028.
Custom AI Silicon at the Hyperscalers: TPU, Trainium, Maia, MTIA and the OpenAI–Broadcom Chip
- 7Supplier roadmap expandsFinancial
Broadcom Q3: AI revenue $16.7 billion, six XPU customers, Anthropic to be the largest in 2027 with 5 GW of TPU 8i
Broadcom's AI semiconductor revenue more than tripled to $16.7 billion with custom accelerators 73 percent of it, and Hock Tan laid out the customer roadmap: multi-tens of billions of dollars of TPUs a year for Google, 1 GW of Ironwood for Anthropic in 2026 then 5 GW of TPU 8i in 2027 and 10 GW more in 2028, 1.3 GW of Jalapeño for OpenAI in 2027 with a second generation near tape-out and a third in development, and MTIA production shipments for Meta. Guidance: $21.7 billion of AI revenue in Q4, $58 billion for fiscal 2026 and about $115 billion in fiscal 2027.
Broadcom Inc. Investor Relations ·
Q3 AI semiconductor revenue of $16.7 billion grew 221% year-over-year, and 54% quarter-over-quarter,
Why it matters here. Broadcom's revenue and six XPU customers show real commercial activity, but its major 2027-28 capacity projections leave room for the execution risk illustrated next by Microsoft's Braga delay.
Custom AI Silicon at the Hyperscalers: TPU, Trainium, Maia, MTIA and the OpenAI–Broadcom Chip
What pushes back
- 8Braga is delayed
Microsoft delays its next Maia chip, Braga, to 2026 after design setbacks
Data Center Dynamics, citing The Information, reported that mass production of Microsoft's next-generation Maia chip, code-named Braga and expected to ship as Maia 200, had been pushed to 2026. The delay left Microsoft behind Google and Amazon on custom silicon and was still visible in Maia 200's limited footprint a year later.
Data Center Dynamics ·
Microsoft has delayed the mass production of its next-generation Maia AI chips until 2026.
What it pushes against. Braga's reported design setbacks and delayed mass production illustrate execution risk, while Maia 200 next shows that deployment within Microsoft still did not mean outside-customer availability.
Custom AI Silicon at the Hyperscalers: TPU, Trainium, Maia, MTIA and the OpenAI–Broadcom Chip
- 9Maia customers remain future
Maia 200: Microsoft's 3nm inference chip runs GPT-5.2 and Copilot, with outside customers 'in the future'
Microsoft announced its second-generation accelerator, a TSMC 3nm part it called the most efficient inference system it had deployed, claiming 30 percent better performance per dollar than alternatives, and began installing it in its US Central region. It runs OpenAI's GPT-5.2, Microsoft 365 Copilot, Foundry and the superintelligence team's work; unlike Maia 100 it is meant to reach cloud customers eventually.
CNBC ·
The Maia 200 comes two years after Microsoft said it had developed its first AI chip, the Maia 100, which was never made available for cloud clients to rent.
What it pushes against. Microsoft's statement that outside Maia 200 customers were for the future pushes against treating deployment inside one provider as broad market availability.
Custom AI Silicon at the Hyperscalers: TPU, Trainium, Maia, MTIA and the OpenAI–Broadcom Chip
- 10Crux access remains planned
Google and Blackstone launch TPU cloud venture Crux AI
Data Center Dynamics reported that the Google and Blackstone neocloud offering TPU services launched as Crux AI. Bloomberg reported that Blackstone provided $5 billion and that Crux AI aims to rent Google AI processors to customers in 2027.
Data Center Dynamics ·
The neocloud set up by Google and Blackstone to offer TPU services has officially launched as Crux AI.
What it pushes against. Crux AI's launch and aim to rent Google processors to customers in 2027 provide an emerging external-access route, but the next entry shows another custom-chip program reserving capacity for its developer's own demand.
Custom AI Silicon at the Hyperscalers: TPU, Trainium, Maia, MTIA and the OpenAI–Broadcom Chip
- 11Jalapeño stays in-house
OpenAI says Jalapeño capacity will serve its own compute demand
OpenAI hardware chief Richard Ho said the company would have its hands full providing compute for OpenAI for a long time. The statement concerned Jalapeño.
Tom's Hardware ·
I think that we're going to have our hands full just providing compute for OpenAI for a good long time.
What it pushes against. OpenAI said it would have its hands full providing compute for itself for a long time, so Jalapeño capacity did not establish broad external availability.
Custom AI Silicon at the Hyperscalers: TPU, Trainium, Maia, MTIA and the OpenAI–Broadcom Chip
Changes to this chain
- Proposal #200 accepted (publish:hyperscaler-custom-accelerators): insert after 9: [counter] Crux access remains planned -> hyperscaler-custom-accelerators#1374: Crux AI's launch and aim to rent Google processors to customers in 2027 provide an emerging external-access route, but the next entry shows another custom-chip program reserving capacity for its developer's own demand.; answer: The evidence shows real progress rather than merely announced silicon: Google trained and served Gemini on TPUs, AWS put nearly 500,000 Trainium2 chips into live Project Rainier, and Meta reported hundreds of thousands of deployed MTIAs. Commitments have also reached multiple gigawatts. But those facts do not establish broad access across customers or on-time execution: Microsoft delayed its next Maia chip after design setbacks and described outside Maia 200 customers as a future prospect. Crux AI launched as a Google-Blackstone TPU cloud venture, but aimed to rent Google processors to customers in 2027, while OpenAI said Jalapeño capacity would serve its own compute demand for a long time. Large deployments and supplier roadmaps therefore show momentum, not yet a broadly available substitute for merchant GPUs.
- Proposal #181 accepted (publish:hyperscaler-custom-accelerators): replace step 10: [counter] Jalapeño stays in-house -> hyperscaler-custom-accelerators#1664: OpenAI said it would have its hands full providing compute for itself for a long time, so Jalapeño capacity did not establish broad external availability.; answer: The evidence shows real progress rather than merely announced silicon: Google trained and served Gemini on TPUs, AWS put nearly 500,000 Trainium2 chips into live Project Rainier, and Meta reported hundreds of thousands of deployed MTIAs. Commitments have also reached multiple gigawatts. But those facts do not establish broad access across customers or on-time execution: Microsoft delayed its next Maia chip after design setbacks and described outside Maia 200 customers as a future prospect, while OpenAI said Jalapeño capacity would serve its own compute demand for a long time. Large deployments and supplier roadmaps therefore show momentum, not yet a broadly available substitute for merchant GPUs.
- Proposal #130 accepted (verify:v2): why of step 2 -> Ironwood was Google's first TPU designed specifically for inference and supported 9,216-chip pods; the next entry shows AWS operating another custom accelerator, Trainium2, at large scale for Claude training and serving.; why of step 4 -> Google said Ironwood would be generally available within weeks, while Meta next reported hundreds of thousands of MTIA chips deployed in production.; why of step 5 -> Meta reported hundreds of thousands of MTIA chips deployed in production, while Anthropic's next agreement committed up to 5 GW of Trainium capacity.; why of step 6 -> Anthropic committed up to 5 GW of Trainium capacity, including capacity planned through the end of 2026; Broadcom next reported commercial XPU revenue while laying out additional TPU capacity scheduled for 2027 and 2028.
- Proposal #40 accepted (verify:v1): why of step 1 -> Gemini's TPU training and serving established production use of a hyperscaler accelerator, which Ironwood extended with a TPU designed specifically for inference.; why of step 2 -> Ironwood showed Google's TPU program tailoring large-scale systems to inference, while Rainier next showed another hyperscaler custom accelerator operating at large live scale for training and serving.; why of step 3 -> Rainier's large live Trainium deployment showed that custom-accelerator scale was operational, while Google's next entry still described Ironwood general availability as due only within weeks.; why of step 4 -> Google's planned Ironwood availability added a prospective access claim, while Meta next supplied separate evidence of a large internally deployed custom-chip fleet.; why of step 5 -> Meta's hundreds of thousands of production MTIAs added a third hyperscaler's internal fleet to the evidence, before Anthropic's Trainium agreement next expressed scale as planned multi-gigawatt capacity.; why of step 6 -> Anthropic's up-to-5-GW Trainium commitment documented multi-gigawatt planned capacity; Broadcom's next roadmap similarly distinguished substantial commercial activity from capacity still scheduled for 2027 and later.; why of step 7 -> Broadcom's revenue and six XPU customers show real commercial activity, but its major 2027-28 capacity projections leave room for the execution risk illustrated next by Microsoft's Braga delay.; why of step 8 -> Braga's reported design setbacks and delayed mass production illustrate execution risk, while Maia 200 next shows that deployment within Microsoft still did not mean outside-customer availability.
Timelines in this chain
- Custom AI Silicon at the Hyperscalers: TPU, Trainium, Maia, MTIA and the OpenAI–Broadcom Chip
- AMD Instinct: From MI300X to Helios and the Gigawatt Deals
- HBM Supply Race: SK hynix, Samsung and Micron from HBM3E to HBM4E
- Nvidia Data Center GPUs: Hopper, Blackwell and Vera Rubin
A chain is written by the editor from entries already published on the timelines above, each of which was read against its source before publication. The chain changes when those entries do. Data and terms.