Uber expanded its long-running AWS contract on April 7, 2026, committing to run real-time ride-matching, driver dispatch, and large-scale model training on Amazon’s custom Graviton4 and Trainium3 silicon, a move Reuters was first to report as part of Uber’s broader AWS cloud deal. The deal, jointly disclosed by Uber and Amazon Web Services, is the most operationally consequential customer win Amazon’s chip program has scored since Anthropic agreed to deploy more than one million Trainium accelerators inside Project Rainier. For an industry where Nvidia still controls roughly 63% of the AI accelerator market, a $200 billion mobility platform moving production inference and training workloads onto Amazon-designed chips is a turning point.
The announcement immediately reframed the AI infrastructure race. Trainium3 delivers up to 2.517 MXFP8 petaflops per chip with 144 GB of HBM3e memory and runs at roughly 30-50% of the cost of comparable Nvidia H100 or H200 hardware, according to figures Amazon disclosed at re:Invent and confirmed in the April 7 brief. Uber’s pilot covers Trainium3 for AI training and Graviton4 for the production ride-matching layer that touches billions of trips per quarter. The strategic message landed harder than the spec sheet: Uber, a star reference customer for Oracle’s OCI and a long-standing Google Cloud partner, just chose Amazon’s silicon for the workloads that decide whether your ride arrives in three minutes or thirteen.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Uber Actually Bought on April 7, 2026
The April 7 disclosure described a two-track expansion rather than a single chip purchase. On the production side, Uber is migrating real-time ride-matching, Trip Serving Zones infrastructure for demand-spike handling, arrival-time estimation, and personalized delivery recommendations onto AWS Graviton4. On the AI side, Uber kicked off a Trainium3 pilot to train the models that decide which driver or courier to dispatch and how to rank suggestions for Uber Eats. Neither company published a headline contract value, but the workloads named in the brief are the same systems Uber has historically priced as multi-hundred-million-dollar annual line items in its infrastructure bill.
Kamran Zargahi, vice president of engineering at Uber, framed the decision around throughput and unit economics rather than research bragging rights. Rich Geraffo, vice president and managing director of North America at AWS, paired the announcement with the existing Anthropic, Apple, and OpenAI references on Trainium. The cumulative customer logo set is what makes the Uber win matter: Anthropic committed to over one million Trainium chips for Project Rainier, OpenAI included Trainium capacity in its $50 billion AWS commitment despite the Microsoft Azure relationship, and Apple publicly praised Trainium performance for Siri-tier inference earlier in 2026. Uber adds the consumer transactional flywheel that none of the other reference customers contribute.
The Trainium3 Spec Sheet: 2.517 PFLOPS and 144 GB HBM3e
Amazon disclosed Trainium3 at re:Invent in November 2025, built on TSMC’s 3nm process as what AWS calls its fourth-generation AI chip platform and first 3nm accelerator, began shipping production silicon in early 2026 as TechTimes reported, and by March 2026 AWS said the chip had already attracted significant capacity commitments from customers across multiple industries. Each Trainium3 package delivers up to 2.517 petaflops in MXFP8 precision, paired with 144 GB of HBM3e memory and 4.9 TB/s of memory bandwidth. The flagship configuration ships as the Trn3 Gen2 UltraServer, which AWS reaffirmed in July 2026 networks up to 144 accelerators behind NeuronLink interconnect for up to 362 MXFP8 petaflops of aggregate compute and 20.7 TB of pooled HBM3e memory. The NL72x2 switched topology delivers approximately 706 TB/s of aggregate bandwidth, putting it ahead of Nvidia’s GB200 NVL72 on raw fabric throughput, per SemiAnalysis Q1 2026 modeling.
The generational jump matters as much as the absolute numbers. At the UltraServer level, AWS reported in July 2026 that Trn3 configurations deliver up to 4.4x the performance and 3.9x the memory bandwidth of comparable Trn2 UltraServers, while landing at roughly four times better performance per watt, a curve Amazon needs because its hyperscale customers are now power-limited rather than capex-limited; by August 2026, AWS had simplified that pitch on its own Trainium product page, framing the chip-to-chip comparison as roughly 2x the compute of Trainium2 alongside the same 4.9 TB/s of memory bandwidth. Trainium3 is fabricated on a custom 3nm-class process at TSMC with CoWoS-L packaging, the same advanced node Nvidia uses for Blackwell. The package draws roughly 750W TDP, slightly above Nvidia’s H200 at 700W and meaningfully below Blackwell B200 at 1,000W. AWS quoted internal customer pricing equivalent to roughly $1.80 per chip-hour for Trainium3 in early 2026, against on-demand H200 pricing closer to $4.80 per chip-hour, a gap TechCrunch sized in March 2026 as up to 50% lower cost than comparable classic cloud servers.
Trainium3 vs Nvidia H200 vs Blackwell B200: Side-by-Side
| Metric | AWS Trainium3 | AWS Trainium2 | Nvidia H200 SXM | Nvidia B200 HGX |
|---|---|---|---|---|
| FP8 / MXFP8 PFLOPS (per chip) | 2.517 | ~1.2 | 1.98 | 4.5 |
| FP4 PFLOPS (per chip) | 2.517 (MXFP4) | None | None | 9.0 |
| Memory capacity | 144 GB HBM3e | 96 GB HBM3 | 141 GB HBM3e | 192 GB HBM3e |
| Memory bandwidth | 4.9 TB/s | 2.9 TB/s | 4.8 TB/s | 8.0 TB/s |
| Scale-up interconnect | NeuronLink (705.6 TB/s/rack) | NeuronLink (~408 TB/s/rack) | NVLink 900 GB/s | NVLink 1.8 TB/s |
| Process node | 3nm-class TSMC, CoWoS-L | 5nm TSMC | TSMC 4NP | TSMC 4NP (dual-die) |
| TDP per package | ~750W | 650W | 700W | 1,000W |
| Cluster size (single fabric) | 144 chips (Trn3 UltraServer) | 64 chips | 72 chips (NVL72) | 72 chips (GB200 NVL72) |
| Approx hourly price | $1.80 (AWS internal) | $1.20 | $4.80 on-demand | $2.12 spot / $6.00 on-demand |
The headline takeaway is that Trainium3 trails Blackwell B200 on raw FP4 and FP8 dense throughput but matches or beats Nvidia H200 on every metric except FP8 dense flops, and undercuts both on price-per-chip-hour. The cluster economics flip even further in Amazon’s favor because NeuronLink scales 144 chips into a single fabric versus 72 for Nvidia’s NVL72 reference rack. For a customer like Uber that runs many concurrent training jobs at sub-frontier scale, the wider single-image fabric and lower hourly rate matter more than the absolute B200 ceiling.
Why Uber Walked Away From Oracle and Hedged Google Cloud
The April 7 deal is, in TechCrunch’s framing, a “thumb-of-the-nose” at Oracle. Uber’s 2023 multi-cloud pact with Oracle Cloud Infrastructure was one of the largest reference deals OCI used to legitimize its hyperscale ambitions, and Uber’s mobility workloads were a frequent talking point in Larry Ellison’s earnings calls. Uber’s decision to migrate its real-time matching layer onto AWS Graviton4 directly rebuts the narrative that OCI’s RDMA fabric and per-core pricing make it the obvious choice for latency-sensitive transactional workloads. It also softens the strategic Google Cloud commitment Uber announced in 2023 by routing the most differentiated AI training pilots through AWS instead of Vertex AI.
Three factors explain the pivot. First, Trainium3’s price-per-petaflop is materially lower than equivalent Nvidia capacity available on Google Cloud or OCI, and Graviton4 already runs Uber’s general-purpose fleet at lower cost than equivalent x86 instances. Second, Amazon committed engineering muscle to Uber’s specific bottlenecks, including custom NeuronLink topology tuning for the recommendation models that drive Uber Eats personalization. Third, Uber wanted optionality away from a single Nvidia-exposed supplier as Blackwell allocations remain rationed and B200 spot pricing fluctuates with hyperscaler demand. AWS Trainium gives Uber a vertically integrated alternative whose roadmap Amazon controls.
The Customer Roster: Anthropic, Apple, OpenAI, Now Uber
Trainium’s customer list looked thin until late 2025. Anthropic’s Project Rainier commitment, structured around Amazon’s original $5 billion follow-on investment and access to over one million Trainium chips, gave the program its anchor workload — a relationship that Anthropic then expanded in an April 2026 deal into a ten-year, roughly $100 billion Trainium compute commitment covering up to 5 gigawatts of capacity. Apple’s February 2026 confirmation that it had moved Siri-tier inference onto Trainium2 added the consumer-grade reliability stamp. OpenAI’s $50 billion AWS commitment, disclosed alongside the broader Bedrock pact that ended Microsoft’s Azure exclusivity, included a defined Trainium capacity carve-out despite the company’s deep CUDA dependencies, and Reuters reported in April 2026 that the pact specifically covers 2 gigawatts of computing power running on Amazon’s Trainium chips. Uber is the fourth marquee logo and the first non-AI-lab transactional workload on the list.
| Customer | Trainium Commitment | Workload | Disclosed |
|---|---|---|---|
| Anthropic | 1M+ chips, Project Rainier | Claude frontier model training | Q4 2025 (expanded Q1 2026) |
| OpenAI | Carve-out within $50B AWS pact | Fine-tuning, inference | Late 2025 |
| Apple | Undisclosed scale | Siri-tier inference, internal training | February 2026 |
| Uber | Pilot scale (expanding) | Ride-matching ML, Eats recommendations | April 7, 2026 |
| Stripe | Inferentia + Trainium2 | Fraud detection inference | Q3 2025 |
| Snap | Trainium2 production | Ranking + content moderation | Q2 2025 |
Andy Jassy’s Multibillion-Dollar Trainium Business
Amazon CEO Andy Jassy first described Trainium as “already a multibillion-dollar business” on the December 2025 earnings cycle, a comment that quietly reset Wall Street’s model for AWS gross margin. AWS reported $28.6 billion in Q1 2026 revenue with operating income up 19% year over year, and analysts now estimate Trainium and Inferentia together drove roughly 40% of the segment’s incremental growth; a September 2026 report from Revenanas pegged AWS’s custom-silicon business, Trainium included, at more than $25 billion in annualized run rate by the second quarter of 2026. The Uber announcement landed three weeks before Amazon’s Q1 earnings call and was almost certainly timed to give Jassy a fresh customer logo to point at. The same weekend, Bloomberg reported Anthropic was exploring a new funding round at over $900 billion in valuation, an outcome that would directly amplify Trainium consumption.
The financial flywheel has three components. First, every dollar Anthropic raises gets reinvested into AWS compute capacity priced in Trainium hours, sending revenue back to Amazon. Second, Uber and similar transactional logos validate Trainium for non-AI-lab buyers, expanding the addressable market beyond frontier-model training. Third, AWS’s vertical integration on the chip stack lets Jassy capture margin Nvidia would otherwise charge, while still selling H200 and B200 capacity through the Bedrock and SageMaker stacks. The result is an AI infrastructure business that compounds whether or not Trainium beats Nvidia on benchmarks, because AWS books the gross margin on whichever silicon customers choose.
What Industry Analysts and Engineers Are Saying
The April 7 deal triggered an unusually candid round of analyst commentary because it crossed the line from AI-lab adoption to mainstream production. “Trainium3 delivers 2x H100 perf at 50% cost, powering Anthropic’s next models,” Andy Jassy said in his re:Invent 2025 keynote, a number AWS has continued to repeat through Q1 2026. Dylan Patel of SemiAnalysis, in his Trainium3 deep dive, called the NL72x2 switched topology “a potential challenger approaching” Nvidia’s NVL72 reference rack on aggregate bandwidth. Tom’s Hardware noted that Trainium3’s 2,517 MXFP8 TFLOPS per package is “nearly two times lower compared to Nvidia’s Blackwell Ultra,” but acknowledged the cluster-level economics close most of the gap.
Kamran Zargahi of Uber framed the choice in operational terms rather than benchmark wars. “Improving Uber rides at scale with AWS Trainium chips” was the headline of the joint Amazon brief, with Zargahi quoted twice on cost-per-inference and latency improvements. Rich Geraffo of AWS positioned the deal as evidence that Trainium had crossed the chasm from research lab to production transaction systems. Independent analysts at Synergy Research pegged AWS at 31% of cloud AI accelerator infrastructure share in Q1 2026, against Nvidia’s 63% silicon share, with the AWS internal Trainium share inside its own fleet now around 12% and rising.
Graviton4: The Unsexy Engine Behind Real-Time Ride-Matching
While Trainium3 captured the headlines, Graviton4 may be the more consequential half of the deal. Graviton4 is Amazon’s fourth-generation Arm-based server CPU, manufactured on TSMC 4nm and shipping with up to 96 cores per socket. AWS publishes 30% better compute performance, 40% better performance on database workloads, and 75% more memory bandwidth than Graviton3. Uber is moving real-time ride-matching, Trip Serving Zones, and surge-handling infrastructure onto Graviton4 instances. These are not exotic AI workloads. They are the latency-bound transactional engines that decide whether the platform clears its target service-level agreement for driver-rider matching during rush hour.
The Arm transition reflects a broader fleet migration that AWS has been pushing for five years. Graviton-based instances now account for over 50% of new CPU capacity at AWS, and the company has steadily shipped Arm-native versions of nearly every managed service. For Uber, the Graviton4 piece reduces the per-trip CPU cost by an estimated 20-25% versus equivalent x86 instances, while delivering predictable latency for the matching algorithm. Combined with the Trainium3 training pipeline, the deal effectively moves Uber’s ride-economics layer onto Amazon-designed silicon end to end.
Project Rainier and the Anthropic Anchor
None of this works without Project Rainier, which TechCrunch reported went live with 500,000 Trainium2 chips in late 2025 before scaling further. Amazon’s original $5 billion follow-on investment in Anthropic, structured around access to more than one million Trainium chips and a multi-gigawatt buildout, has since scaled into a ten-year, roughly $100 billion Trainium compute commitment that Anthropic secured in April 2026, giving it access to up to 5 gigawatts of capacity. That commitment is the foundational customer commitment that justifies Trainium’s full production volume: TechCrunch reported in March 2026 that Anthropic’s Claude was already running on more than 1 million Trainium2 chips, part of a fleet AWS had grown to 1.4 million Trainium chips deployed across three generations, and Reuters separately reported that same month that Anthropic expects roughly 1 gigawatt of combined Trainium2/Trainium3 capacity to be live by the end of the year. Anthropic’s Claude Opus 4.6 and Sonnet 4.6 training runs are the workloads that have stress-tested NeuronLink, the Neuron compiler stack, and the multi-tenant fabric. Without Anthropic, Trainium would still be a sub-scale science project. With Anthropic anchoring the production tape-out, Amazon can amortize the chip’s NRE cost across enough volume to justify the 3nm-class node and CoWoS-L packaging.
The April 7 Uber deal is the second-order win that Project Rainier was supposed to enable. Once the foundational anchor proves the chip’s reliability, smaller transactional customers can adopt Trainium without taking on first-mover risk. Apple validated the consumer reliability story in February. Uber now validates the latency-sensitive transactional story. The next batch of references AWS will likely surface includes large e-commerce, SaaS observability, and ad-tech ranking customers who have been quietly piloting Trainium2 since 2025. Each new logo gives the next prospect cover to switch.
Competitive Fallout: Oracle, Google Cloud, and Nvidia
The biggest immediate loser is Oracle. Uber was a flagship reference for OCI’s Generation 2 cloud, and the public migration of ride-matching workloads to AWS Graviton4 cuts directly against Oracle’s positioning as the latency-optimized hyperscaler. Larry Ellison’s 2025 commentary on OCI’s $300 billion backlog included repeated references to Uber-class workloads, and analysts will press management on the April 7 disclosure during the next earnings cycle. Google Cloud comes out as a partial loser too, since the Vertex AI pitch for Uber’s recommendation models did not win the new training pilot. The deal does not unwind Uber’s existing Google Cloud commitments, but it caps the share-of-wallet expansion path Google had penciled in.
Nvidia is the strategic loser even though it remains the volume leader. Every workload that ships on Trainium3 instead of H200 or B200 is revenue Nvidia does not book, and the Uber logo gives every AWS sales rep a credible reference for transactional buyers who previously assumed CUDA was non-negotiable. Nvidia’s response is the Vera Rubin platform, which combines the Vera CPU with Rubin R100 GPUs in 336-billion-transistor packages and targets a 5x leap over Blackwell on training throughput. The Rubin ramp is timed to head off exactly the kind of price erosion AWS is now driving, but Trainium3’s price advantage compounds for at least another four quarters before Rubin reaches volume.
Historical Context: From Trainium1 to a Multibillion-Dollar Business
Trainium’s story tracks AWS’s broader patience with custom silicon. Trainium1 launched in late 2021 to limited fanfare, with most analysts dismissing it as a defensive in-house tool rather than a credible Nvidia competitor. Trainium2 arrived in late 2024 with enough capacity and software maturity to attract Anthropic, but adoption beyond AWS-native ML teams was slow; as of August 2026, AWS still lists its Trn2 UltraServer configuration — which pools 64 Trainium2 chips for 83.2 FP8 petaflops, 6 TB of total HBM, and 12.8 Tbps of EFAv3 networking — as a preview offering even as Trainium3 has moved to general availability. AWS’s own EC2 documentation, current as of August 2026, still credits Trn2-based instances with 30-40% better price-performance than comparable Nvidia-based P5e and P5en instances, the figure Amazon leans on whenever customers ask why they should evaluate an older Trainium generation instead of jumping straight to Trainium3. Trainium3, launched April 2025 in initial sampling and entering production by Q1 2026, is the first generation with the bandwidth, memory, and software ecosystem to support frontier-class training on par with Nvidia clusters. Each generation roughly doubled per-chip throughput, while AWS quietly built the Neuron SDK, Bedrock model marketplace, and SageMaker training stack to make migration feasible for non-research customers.
Inferentia, Trainium’s smaller-volume sibling, deserves credit too. Inferentia2 has been quietly handling production inference for Stripe fraud models, Snap content moderation, and AWS-native services for years, building the operational track record that makes a Uber-scale customer comfortable with Trainium. The two chip families now cover the full ML lifecycle inside AWS, and the company is preparing a Trainium4 design tape-out for late 2026 that will move to a 2nm-class process and add native MXFP4 support across the cluster fabric. The roadmap is the answer to the question of whether Trainium is a one-generation phenomenon: Amazon is committed to a multi-decade silicon arc.
Five Predictions for Trainium and the AI Chip Race
First, Trainium will cross 25% internal AWS AI accelerator share by Q4 2026. The current trajectory of 12% and rising, plus the Uber and Apple validation, plus Anthropic’s Project Rainier ramp, makes the curve inevitable. Second, Nvidia will respond with aggressive Blackwell B200 price cuts targeted at AWS-competitive accounts, particularly transactional customers who could otherwise migrate. Watch for B200 spot pricing to settle below $1.80 per chip-hour in non-allocated regions by mid-2026. Third, at least two more Fortune 100 logos beyond Uber and Apple will publicly disclose Trainium pilots before re:Invent 2026, with retail and financial services as the most likely categories.
Fourth, Google will respond by accelerating TPU v8t and v8i availability beyond the Anthropic-only commitment that drove its $40 billion follow-on investment. Public Vertex AI access to TPU v8 capacity is the only structural counter to Trainium’s transactional story, and Google’s $21 billion Broadcom partnership is sized to support exactly that scale-out. Fifth, Microsoft Azure remains the wildcard. Azure has no in-house AI chip at Trainium’s scale, and the OpenAI Bedrock pact already eroded its exclusivity moat. Expect Microsoft to announce its own custom AI accelerator generation by re:Invent 2026 or to deepen its co-design relationship with AMD’s MI400 series as a hedge.
Market Impact: Stocks, Capex, and the Power Constraint
The April 7 announcement did not produce a dramatic single-day stock move, but it landed in a market already grappling with Big Tech’s $635 billion 2026 AI infrastructure capex commitment and the energy constraints flagged by Morgan Stanley and S&P Global. Amazon’s stock traded modestly higher in the days after the disclosure, while Oracle drifted on the relative-positioning concern. The deal’s bigger market impact is the second-order signal it sends to enterprise procurement teams: AWS Trainium is no longer experimental, and CIOs who renew Nvidia-heavy infrastructure agreements without evaluating Trainium are leaving cost savings on the table.
Power is the binding constraint that makes Trainium’s energy efficiency more valuable than its raw FLOPS. Trainium3’s 4x energy-efficiency improvement over Trainium2 directly lowers the megawatt cost of every training job, and at hyperscale that translates into more usable compute per gigawatt of contracted power. Half of U.S. AI data centers planned through 2027 are already delayed or canceled due to a 7 GW capacity crisis in the grid, and every chip with better performance-per-watt becomes a strategic asset. AWS’s pitch to Uber, and to every prospective Trainium customer, increasingly leads with kilowatt-hours per inference rather than dollars per chip.
What Could Derail Amazon’s Trainium Momentum
Three risks could blunt the Trainium curve. Software compatibility remains the largest, though it is narrowing: AWS shipped Neuron SDK 2.30.0 in May 2026, adding 22 new NKI kernels and Trainium3-specific capabilities that cut down on custom kernel work for porting teams. Even so, CUDA’s ecosystem advantage is still measured in years, and the Neuron SDK requires customers to either rewrite kernels or rely on Amazon’s compiler to translate PyTorch and JAX workloads. Anthropic’s deep co-design with AWS papered over this gap, but Uber and Apple are different profiles, and any high-profile training run that fails to converge on Trainium will become a sales counter for Nvidia. Second, Blackwell B200 supply normalization in late 2026 could reverse Trainium’s price-per-chip-hour advantage, particularly if Nvidia chooses to compete aggressively on hyperscale rate cards.
Third, the customer concentration risk is real. Anthropic represents an outsized share of Trainium consumption, and any disruption to Anthropic’s funding trajectory or training cadence would expose AWS to underutilized fabs. The $350 billion tender offer Anthropic ran earlier in 2026 fell short of its target, illustrating that even the most well-capitalized AI labs face sentiment cycles. Diversification toward transactional customers like Uber and Apple is precisely how Amazon hedges this risk, but the diversification is still in its early innings. The April 7 deal is a strategic step in the right direction, not a finished customer mix.
Key Takeaways for Cloud Buyers
Enterprise architects evaluating AI infrastructure renewals in 2026 should treat the Uber announcement as a forcing function. Three concrete actions follow from the deal. First, request a Trainium3 evaluation from AWS for any greenfield training workload above $5 million in annual spend, and benchmark against Nvidia H200 and B200 quotes from the same hyperscaler. Second, evaluate Graviton4 for any latency-bound CPU workload that currently runs on x86 instances, with an expected 20-25% unit-cost reduction. Third, build a multi-vendor chip strategy that hedges between Nvidia, Trainium, and Google TPU rather than betting the entire ML stack on one supplier.
The deal also reframes the build-vs-buy debate for AI compute. Until April 2026, the conventional answer was that only AI labs at Anthropic’s scale could justify alternative-silicon evaluations. Uber’s pilot pushes that threshold dramatically lower. If a transactional consumer platform with conservative reliability requirements is willing to pilot Trainium3, then virtually any enterprise with a serious AI roadmap should at least include Trainium in the procurement matrix. The cost of evaluation is now low, the customer references are credible, and the price advantage is meaningful enough to fund an internal benchmarking exercise.
FAQ: AWS Trainium3 and the Uber Deal
What did Uber and Amazon announce on April 7, 2026?
Uber expanded its existing AWS contract to migrate real-time ride-matching, Trip Serving Zones, arrival-time estimation, and personalized delivery recommendations onto AWS Graviton4 CPUs, and to pilot AI training of dispatch and ranking models on AWS Trainium3 accelerators. The companies did not disclose a contract value, but the named workloads represent a multi-hundred-million-dollar annual infrastructure commitment.
How does AWS Trainium3 compare to Nvidia H200 and B200?
Trainium3 delivers 2.517 MXFP8 PFLOPS per chip, 144 GB of HBM3e memory, and 4.9 TB/s of memory bandwidth. That positions it ahead of Nvidia H200 on memory bandwidth and FP8 throughput, and behind Blackwell B200 on raw FP4 and FP8 dense flops. Trainium3 wins on hourly cost-per-chip and on cluster-level fabric bandwidth via NeuronLink: AWS confirmed in July 2026 that a full 144-chip Trn3 Gen2 UltraServer scales to roughly 362 MXFP8 petaflops, 20.7 TB of pooled HBM3e, and 706 TB/s of aggregate memory bandwidth, dwarfing the 64-chip Trn2 UltraServer’s 83.2 FP8 petaflops and 6 TB of HBM.
How much cheaper is Trainium3 than equivalent Nvidia hardware?
AWS quotes Trainium3 at roughly 30-50% of the cost of comparable Nvidia H100 or H200 capacity. Internal AWS pricing equates to approximately $1.80 per chip-hour for Trainium3, against on-demand H200 pricing closer to $4.80 per chip-hour and B200 on-demand pricing around $6.00 per chip-hour. Spot pricing closes some of the gap, but Trainium3 remains the lower-cost option in nearly every comparable configuration.
Which other major customers run on AWS Trainium?
Anthropic committed to over one million Trainium chips for Project Rainier, OpenAI carved out Trainium capacity within its $50 billion AWS commitment, and Apple confirmed in February 2026 that it had moved Siri-tier inference workloads to Trainium2. Stripe and Snap have run production Inferentia and Trainium2 workloads for fraud detection and content ranking since 2025. Uber is now the latest and largest transactional reference.
What is AWS Project Rainier?
Project Rainier is the joint AWS and Anthropic infrastructure buildout that combines a multi-gigawatt power footprint with access to over one million Trainium accelerators for Anthropic’s frontier model training. Amazon committed an additional $5 billion to Anthropic to anchor the project, an investment that grew into a ten-year, roughly $100 billion Trainium compute commitment in April 2026 covering up to 5 gigawatts of capacity, and the buildout now serves as the production proving ground for Trainium2 and Trainium3 reliability, scaling, and software stack maturity.
What is the strategic fallout for Oracle and Google Cloud?
Oracle loses its highest-profile mobility reference for Generation 2 OCI, since Uber chose AWS Graviton4 over OCI for the latency-bound ride-matching workload that Larry Ellison frequently cited. Google Cloud retains its existing Uber commitments but loses the share-of-wallet expansion path it had penciled in for Vertex AI training. Both clouds will face investor questions about reference-customer momentum on the next earnings cycle.
Will the Trainium3 deal materially change AWS revenue?
Andy Jassy already described Trainium as a multibillion-dollar business in late 2025, and Trainium3 plus Inferentia together are estimated to drive roughly 40% of AWS’s Q1 2026 incremental segment growth. The Uber deal alone will not move the needle on AWS’s $28.6 billion quarterly revenue, but it is part of a broader customer-mix shift that supports continued share-take in the AI accelerator market over the next four quarters.
Where can I read AWS’s official statement?
AWS published its joint announcement with Uber on the About Amazon newsroom, with technical details on Trainium3 specifications and the Trn3 Gen2 UltraServer covered in Tom’s Hardware and the SemiAnalysis Trainium3 deep dive. TechCrunch’s coverage of the Uber deal and the official AWS Trainium product page round out the primary sources.
