Video summary: this article is a commented summary — and checked — from the video “No Nvidia Chips Needed! Amazon's New AI Data Center For Anthropic Is Truly Massive”, from the channel CNBC, published on 29 October 2025.
Video URL: https://www.youtube.com/watch?v=vnGC4YS36gU
In October 2025 the CNBC entered, with camera, into what would become the largest cluster of AI chips non-NVIDIA in the world: the Project Rainier, from AWS, in New Carlisle, Indiana. One year before that was a cornfield. Today there are seven buildings full of Trainium2 designed by Amazon itself, running exclusively for a single customer: Anthropic, maker of Claude.
The video is good, but it's almost a year old. Besides summarizing what was shown, this post checks claim by claim — what was confirmed, what changed, and what has aged poorly since then.
The site numbers: 1.200 acres, 30 buildings, 2,2 GW
The land has 1.200 acres (about 486 hectares). In the recording, seven buildings were in full operation; the complete project plans for 30 buildings across three campuses, consuming about 2,2 gigawatts of electricity — energy equivalent to that of more than 1,5 million homes in the local utility's area. Amazon estimates at least two more years of construction to complete the complex.
The chronology shows the urgency:
- End of 2022 — around the time ChatGPT launched, Amazon begins looking for land. American Electric Power offers New Carlisle due to the abundance of extra-high voltage transmission lines and a robust substation.
- Spring of 2023 -- first visits to the site.
- 2024 -- agreement of US$ 11 billions officialized: the largest capital investment in Indiana's history.
- September 2024 -- start of construction.
- October 2025 -- seven buildings in full operation. From zero to supercomputer in ~13 months.
None of this came for free for the public sector: the county granted more than US$ 4 billions in property tax exemptions and technology equipment tax exemptions over 35 years, and a state law of 2019 should save another ~US$ 4 billions in 50 years. In exchange, Indiana talks about more than US$ 1 billion of GDP increment and ~9.000 jobs -- all temporary, of construction. The promised permanents are 1.000, with at least 600 above the county average salary.
Trainium: why AWS removed NVIDIA from the game
Inside the buildings there is not a single NVIDIA GPU. Two towers form an UltraServer, and each UltraServer carries 64 Trainium2 chips. Amazon has been designing silicon since 2013 and the lineage is long:
- 2013 — Nitro, the virtualization/offload chip that powers modern EC2.
- 2015 — acquisition of the Israeli Annapurna Labs, the team that came to design everything.
- 2018 — Graviton, the company's ARM server processor.
- 2019 — Inferentia, the first AI (inference) chip.
- 2021 — Trainium1, admittedly a “learning generation”.
- December of 2024 — Trainium2, the chip that fills the Project Rainier.
The alleged advantage isn't raw performance: it's cost and availability. The Trainium is simpler and less powerful than a top-tier GPU, but AWS puts twice as many of them in each building and keeps power and cooling under control. And, as an analyst summarized in the video, you don't need to “wait in line for Jensen to deliver what you need”.
If you want to understand the architectural difference between GPUs and dedicated AI chips, we have two posts that cover exactly that: GPU vs TPU: how AI chips work and Google TPUs: architecture, generations, and models.
Anthropic: sole tenant, multi-cloud strategy
The Project Rainier is dedicated to a single client. Anthropic, founded in 2021 by former OpenAI researchers, has used AWS as its primary cloud since Amazon's first major investment, in 2023 — at the time of the video, about US$ 8 billion. The two companies co-developed the chip: Anthropic's feedback on what the model needs went into the Trainium3 project.
But Anthropic isn't monogamous. Days before the report it closed with Google the largest TPU contract in its history — up to 1 million TPUs and more than 1 GW of capacity in 2026. In the words of CPO Mike Krieger, the multi-chip strategy was the only way to meet demand: “we take the capacity we can, where we can”.
When asked if Amazon has ever thought about buying Anthropic, the AWS CEO, Matt Garman, was direct: “no. We love the partnership as it is”.
Energy and water: the bill left for New Carlisle
New Carlisle has 1.900 residents. It receives daily about 4.000 workers construction workers — and, beside it, an electric car battery plant worth US$ 3,5 billion, from GM in partnership with Samsung. The highways 20 and 2 became dangerous; Amazon is paying US$ 7 million in road improvements (negotiating up to US$ 15 million more) and US$ 114 million for a water treatment plant and sewage system.
The electricity bill is the most sensitive point. Indiana exempts large data centers from the state tax on 7% on electricity. The utility I&M projects that peak demand more than doubles — from ~2,8 GW in 2024 to more than 7 GW by 2030 — and admits that the two data centers it will serve (Amazon and a Google one, worth US$ 2 billion, in Fort Wayne) account for nearly all the electric load it serves today. A survey cited in the report points to monthly bills up to 267% larger in areas near new data centers compared to five years ago.
The energy mix also weighs heavily: 53% of I&M's power comes from a nearby nuclear plant and 35% from a coal-fired thermal plant scheduled for closure in 2028 — a timeline the Trump administration signaled it would delay for most coal plants in the country, precisely because of AI. Regarding the net zero target in 2040, Garman was honest: “The path from here to 2040 won't be linear. Will there be a need for gas generation? I think absolutely yes.”
In water, AWS claims to cool the site with outside air 98% of the year, using water only for the remaining 2%, and reducing consumption by 23% per year with on-site treatment — but they won't disclose the absolute gallon figure. The municipality is still studying local aquifers to see if they can support both Amazon and GM together.
What was confirmed (and what changed) since October 2025
Here's the part the video couldn't have. Each line below was verified against official announcements and subsequent coverage:
| Claim in the video (out/2025) | Verified status |
|---|---|
| ~500 thousand Trainium2 chips running in Indiana | Confirmed. AWS officially announced the activation of Project Rainier in 29/10/2025 with this volume. |
| “Over 1 million Trainium2 by the end of 2025” | Confirmed. The Anthropic fleet exceeded 1 million Trainium2; adding all customers, AWS reported ~1,4 million Trainium chips deployed across three generations. |
| 1.200 acres, 30 buildings, 2,2 GW | Confirmed. And there was more: in 24/11/2025 AWS announced a second set of campuses of ~US$ 15 billion in northern Indiana (NIPSCO territory), with ~2,4 GW of its own — bringing AWS's commitment in the state to ~US$ 26 billion and 4,6 GW. |
| Trainium3 “later this year” | Delivered. Announced with general availability at re:Invent, in 2/12/2025: 3 nm from TSMC, 144 GB of HBM3e, ~2,5 PFLOPS FP8 per chip and UltraServers of up to 144 chips — which deliver up to 4,4x more performance, 4x more energy efficiency and almost 4x more memory bandwidth than Trainium2 UltraServers. It's already being deployed in New Carlisle. The Trainium4 was previewed in the same keynote, with NVIDIA NVLink Fusion support and deliveries expected for 2027 — irony aside. |
| Trainium2 with “30 to 40% better cost-benefit” than GPUs | Partially. Andy Jassy started citing ~30% comparable GPUs for Trainium2; the 30–40% range became the metric for Trainium3 over Trainium2. It's not the same number measuring the same thing. |
| US$ 8 billion from Amazon in Anthropic | Correct at the time, outdated today. In 20/04/2026 Amazon invested more $$ 5 billion (total of $$ 13 billion) and promised up to $$ 20 billion additional tied to commercial targets. In return, Anthropic committed to spending more than $$ 100 billion on AWS over ten years and reserved up to 5 GW of Trainium capacity. |
| Anthropic secured up to 1 million TPUs from Google | Confirmed (23/10/2025 announcement), with more than 1 GW coming online at 2026. |
| Garman: “I'm not discounting the checks OpenAI is writing” | Aged poorly. In 03/11/2025 — five days after the video aired — AWS signed a $$ 38 billion seven-year contract with OpenAI. Then came the expansion, including ~2 GW of Trainium. |
| I&M peak demand doubling from 2,8 GW to 7+ GW by 2030 | Confirmed in the utility's own projection. |
| Air cooling 98% of the year | Confirmed. AWS maintains the number: water only in ~2% of operating hours, about a week per year. The water context also became clearer — the safe yield estimated of the Kankakee aquifer is ~44 million gallons/day, and the two New Carlisle treatment plants can deliver up to 24 million/day for all industrial use; Amazon is permitted a fraction of that. |
Two relevant things happened depois and deserve to go on the record:
- 27 of 2026 — the load cut. During a winter storm, the US Department of Energy ordered the site to cut 50 MW from the grid. Emergency diesel generators came online, but the cold froze pneumatic lines in five of them, compromising emissions control for five hours and releasing pollutants linked to smog and acid rain. The case only became public because Amazon needed to report it to Indiana's environmental agency. It's the practical demonstration of what residents feared in the video: the network is the bottleneck, and the data center is the first load to be cut.
- The chip business turned into real revenue. On 2026 Amazon reported that its custom silicon line exceeded US$ 20 billion of annualized revenue, with more than US$ 225 billion in multi-year Trainium commitments — Anthropic, OpenAI, and clients like Uber. The Trainium3 was practically sold out before it even existed in volume.
What this means in practice
Three takeaways that matter for those working with infrastructure:
1. NVIDIA's de facto monopoly in training has real cracks. Not because Trainium is better — it isn't, chip for chip — but because “good enough, available now and cheaper” wins contracts. Google (TPU), Amazon (Trainium), and now Anthropic buying from both prove the market wants a second source.
2. The constraint became energy, not silicon. The entire bottleneck of this report is electrical: transmission lines, substation, electricity tax exemption, coal that doesn't close, gas that comes in, cut of 50 MW in a snowstorm. Who projects capacity in 2026 sizes megawatts before sizing racks.
3. The social cost is local and the revenue is global. A city of 1.900 people absorbs traffic, water and risk of brownout while the value generated is accounted for in Seattle. As one analyst summarized: “it's going to be a constant fight for data center builders, because by the amount of land and resources they use, they don't generate that many jobs”.
If you want the background on how Amazon got here, it's worth The story of Jeff Bezos: from the garage to the cloud. And if the idea is to run AI on your own instead of renting 2,2 GW from AWS, start with Running local AI on Linux with Ollama or by refurbished AI server with 64 GB of VRAM.