TL;DR
On October 27, 2026, Oracle unified NVIDIA Vera Rubin NVL72 GA, 2027 Google TPU launch, 2027 Cerebras inference and SambaNova SN50 RDU into one open multi-vendor cloud — while AWS Trainium stays locked to EC2 and the Neuron SDK, Microsoft Maia 300 plans gigawatt-scale custom deployment, and Google TPU spills out to OCI only in 2027.
Comparison Overview
On October 27, 2026, at Oracle AI World, Oracle announced on a single day: NVIDIA Vera Rubin NVL72 reaching GA on OCI, Google TPUs arriving in 2027, and Cerebras inference arriving in 2027; per IT Home, SambaNova SN50 RDU also joined OCI the same day, and OCI has already brought Positron Atlas live alongside NVIDIA and AMD. Oracle thus assembled four to five chip vendors in a multi-vendor model. Meanwhile, AWS Trainium, Microsoft Maia and Google TPU follow in-house silicon routes. This article compares the four cloud vendors across custom chips, third-party chip supply, cross-cloud availability, anchor customers and current scale milestones.
Comparison Table
| Dimension | Oracle OCI | AWS | Microsoft | Google Cloud |
|---|---|---|---|---|
| In-house AI chip | No in-house AI chip; multi-vendor route | Trainium: Trainium 3 delivers 4x+ performance vs previous generation with significantly lower energy use | Maia 300: launched October 12, 2026; ~300,000 chips at TSMC; gigawatt-scale deployment planned | TPU: v7 Ironwood reached GCP GA in March 2026; ~4.6 PFLOPS FP8, 192 GB HBM3e, 7.37 TB/s bandwidth |
| Third-party chips offered | NVIDIA Vera Rubin NVL72 (GA October 27, 2026), Google TPU from 2027, Cerebras inference from 2027, SambaNova SN50 RDU, Positron Atlas already live | Not disclosed (AWS third-party chip supply not listed in this research) | Not disclosed | Not disclosed (third-party chips offered by Google Cloud not listed in this research) |
| Cross-cloud chip portability | Customers can choose NVIDIA GPUs and Google TPUs within the same OCI cloud | Trainium is available only through AWS EC2 and tied to the Neuron SDK (per Spheron analysis, October 19, 2026) | Not disclosed (Maia 300 cross-cloud scope not listed in this research) | External TPU revenue began recognition in the fiscal quarter ended June 2026; from 2027 TPUs are offered via OCI, the first hyperscaler outside Google |
| Anchor customers | Not listed (no OCI chip customer roster within this research) | Anthropic, OpenAI and Decart use Trainium (per Amazon's official explainer) | Not listed | Anthropic committed to scaling Google TPU usage to as many as 1 million chips, with 1+ GW online in 2026 |
| Current scale milestone | Vera Rubin NVL72 GA on October 27, 2026, scaling to gigawatt levels | Trainium 3 launched, 4x+ performance vs previous generation | Maia 300 launched October 12, 2026; ~300,000 chips at TSMC | TPU v7 Ironwood reached GCP GA in March 2026 |
Evidence and Positions
Oracle's official line (Oracle press release, October 27, 2026): Vera Rubin NVL72 and Vera CPU reached GA on OCI and scale to gigawatt levels; OCI will be the first hyperscaler outside Google to offer Google TPUs, launching in calendar 2027; Cerebras high-speed inference launches on OCI in calendar 2027. Per IT Home, SambaNova SN50 RDU also joined OCI the same day, and OCI has already brought Positron Atlas live beyond NVIDIA and AMD.
AWS's official line (Amazon's explainer): Trainium 3 delivers 4x+ performance versus the previous generation with significantly lower energy use; Anthropic, OpenAI and Decart use Trainium. Per Spheron's analysis (October 19, 2026), Trainium is available only through AWS EC2 and is tied to the Neuron SDK.
Microsoft and Google disclosures: per EET-China / shattered.io reporting (October 2026), Microsoft Maia 300 launched on October 12, 2026 with about 300,000 chips at TSMC and gigawatt-scale deployment planned; Google TPU v7 Ironwood reached GCP GA in March 2026 at ~4.6 PFLOPS FP8, 192 GB HBM3e and 7.37 TB/s, with external TPU revenue recognized from the fiscal quarter ended June 2026; Anthropic committed to scaling TPU usage to as many as 1 million chips with 1+ GW online in 2026.
Interpreting the Disagreement
First, an openness spectrum: Oracle builds no in-house chip and packs NVIDIA, Google TPU, Cerebras, SambaNova and Positron into one cloud; AWS locks Trainium inside EC2 and the Neuron SDK (per Spheron); Microsoft Maia 300 and Google TPU mainly serve workloads inside their own clouds, with Google TPU spilling out to OCI only in 2027. Openness runs from OCI down toward AWS.
Second, performance figures are not comparable: AWS gives Trainium 3 only a relative figure (4x+ vs previous generation, lower energy) with no absolute number; Google gives absolute specs for TPU v7 Ironwood (~4.6 PFLOPS FP8, 192 GB HBM3e); Microsoft gives a shipment-capacity figure for Maia 300 (~300,000 chips at TSMC). Relative gain, absolute spec and foundry shipment volume cannot be converted across each other.
Third, customer rosters are asymmetric: AWS officially names Anthropic, OpenAI and Decart on Trainium; Google's side has Anthropic's commitment to up to 1 million TPUs with 1+ GW online in 2026; OCI and Microsoft list no chip customer rosters within this research, so the actual relationship between the open model and customer lock-in can only be partially judged.
Data Sources
- Oracle press release - AI World: Oracle brings Google TPUs to OCI
- IT Home (Toutiao mirror) - SambaNova SN50 RDU joins OCI
- Amazon official - AWS Trainium and Graviton explained
- Spheron - Hyperscaler custom AI chips 2026: Trainium, TPU, Maia, MTIA vs NVIDIA GPU
- EET-China - Report on Google external TPU sales, Anthropic scale-up and Maia 300