SemiAnalysis chief executive Dylan Patel used a talk at Modal’s Runtime 2026 conference to explain why his firm’s ClusterMAX 3.0 rating puts Amazon Web Services in Bronze. “How is AWS bronze?” he asked. “It turns out to like get a cluster, you have to talk to like 10 people, and it takes forever”.
The rating, published on 23 September, gives a medal to 19 of the 77 GPU clouds SemiAnalysis reviewed. The previous edition, ClusterMAX 2.0 in November 2025, gave a medal to 37. By our count, 10 of the 19 medals are now Bronze.
The talk was given on 1 October in San Francisco, according to Modal’s event page, and the video was posted to YouTube on 2 November. Quotes below are from that video, with timestamps. Ratings are quoted from SemiAnalysis’s own ClusterMAX 3.0 report.
The Source Talk · Dylan Patel at Runtime 2026
Dylan Patel, CEO of SemiAnalysis, “Compute, capital, and who actually delivers it”, Modal Runtime 2026, 1 October 2026. The ClusterMAX section runs from about 9:48 to 13:10.
The key insight: SemiAnalysis’s ClusterMAX 3.0 gives a medal to 19 GPU clouds, down from 37 in ClusterMAX 2.0, a 49 per cent drop by our arithmetic. Ten of the 19 are Bronze, by our count, and none of the four largest US clouds in the ratings is in Platinum.
Patel described the scale at [9:54]: there are “323 providers that we’ve tracked”, “77 that we actually reviewed, and 19 that we gave medallions to”. He called “CoreWeave and Nebius” “the best providers in our view” and said “Oracle, Google are really good as well”. Google, he said, was Bronze in the first ClusterMAX and has moved up since.
On the way down, at [12:11], he listed: “Azure got downgraded to silver, Crusoe got downgraded to bronze” … “and AWS got downgraded to bronze”. Together was also moved to Bronze. He said ClusterMAX keeps raising the bar, so “there’ve been a lot of folks who’ve been downgraded over time”.
SemiAnalysis ClusterMAX tiers for managed GPU clusters, ClusterMAX 2.0 (November 2025) against ClusterMAX 3.0 (September 2026), for nine named providers. The 3.0 medal split of 2 Platinum, 2 Gold, 5 Silver and 10 Bronze is our count of the report’s provider reviews. Tiers are SemiAnalysis’s ratings. Sources: SemiAnalysis ClusterMAX 2.0 and 3.0.
The Medal Table
The report’s summary says “Nebius joins CoreWeave in the Platinum tier” and “Google Cloud joins Oracle in the Gold tier”. It adds: “We raise the bar this round as only 19 neoclouds globally achieve a Medallion rating”.
By our count of the report’s provider reviews, the 19 split into 2 Platinum, 2 Gold, 5 Silver and 10 Bronze. Silver is Lambda, Microsoft, Firmus, TensorWave and GMI. Bronze includes AWS, Together and Crusoe alongside seven smaller clouds.
Below Bronze, SemiAnalysis created a new tier it calls the Participation Ribbon. The report says “15 providers join this rating” and that it describes providers that “do the bare minimum to get by”.
By our arithmetic, medal holders fell from 37 in ClusterMAX 2.0 to 19, a 49 per cent drop. The 19 are about one in four of the 77 reviewed and 5.9 per cent of the 323 tracked.
Why AWS Is Bronze
Patel said AWS’s “networking is a little bit difficult to deal with” and that “it’s just like not an easy user experience”. He added: “Their HyperPod offering is not that great”.
The report’s AWS review reaches a short verdict: “Managed clusters are not AWS’s priority”. It says “The difficulty with AWS begins before you get on the cluster”.
It also records strengths. One AWS cluster type caught an injected GPU fault and “gave us a new and healthy node only 24m after the original injection”. The report says “AWS’s specialists and engineers were intelligent and responsive in our interactions with them”.
In ClusterMAX 2.0, SemiAnalysis placed AWS in Silver, with Google and Together: “Google rises to the top of the Silver tier, alongside AWS, together.ai and Lambda”.
Azure, Crusoe and Together
Azure was Gold in November 2025. The 3.0 review opens: “Using Azure clusters can be very difficult due to the Microsoft bureaucracy”. The report’s summary says “Azure moves to Silver”.
Crusoe was Gold in ClusterMAX 2.0. The new report says “Crusoe comes in towards the bottom of our Bronze category for this round”. For Together, which was Silver, the report cites reliability issues as “cause for downgrading it to Bronze”.
The Failure Test
Part of the rating is how fast a cloud notices broken hardware. At [12:42], Patel said SemiAnalysis would “inject failures” that mimic an Nvidia GPU error and time the response.
“And many clouds would not catch it until you alerted them. And many clouds, even once you alerted them, would take days to respond”, he said. In his view a cloud “should catch that in their own monitoring and fix it themselves”.
The Security Finding
Security is now part of the score. At [10:40], Patel said: “With ClusterMax V3, we asked some models to hack the neo clouds”. On quite a few, he said, models could use vulnerabilities that had already been published and “access other customers’ storage”.
In one case, he said, an open-source model gained access across an entire cluster. The team could see that “a certain country’s” … “defense industry including their national security” … “intelligence agency” were on that cluster, along with “their job names”. “We immediately closed out and we’re like, guys, this is ridiculous”, he said.
SemiAnalysis’s August article, Most Neoclouds Suck At Security, describes the same testing. It says exposed data included customer information from banks, telcos, universities, research institutions and AI labs, and “in one case, a national intelligence agency from a country with a top-10 global GDP”. It adds: “We disclosed this immediately, the provider patched it within a week, and we verified the patch”.
The fixes SemiAnalysis recommends are basic: keep drivers and software up to date, join Nvidia’s embargoed security program to get early notice of flaws, and run its free audit tool. The 3.0 report’s security section ends: “Be a good neighbor and don’t get hacked”.
Business Pill · Least Privilege
An evergreen Business Engineer explainer on least-privilege access: giving each user or system only the access its role needs. It teaches the general idea only and does not discuss SemiAnalysis, ClusterMAX or any provider in this story.
As we read the two reports, four of the nine providers in our chart moved down a tier or more, and SemiAnalysis says it raised the bar this round.
As we read Patel’s examples, the tests that moved ratings are operational: how long it takes to get a cluster, whether a cloud catches a fault before the customer reports it, and whether one tenant can see another.
Dylan Patel, CEO, SemiAnalysis, Runtime 2026, 1 October 2026
“And many clouds, even once you alerted them, would take days to respond”
Three Implications
FOR AWS AND AZURE Their ClusterMAX tiers are Bronze and Silver, below two neoclouds. SemiAnalysis’s report says managed clusters are not AWS’s priority.
FOR NEOCLOUD BUYERS SemiAnalysis’s security article says exposed data included banks, telcos, universities, research institutions and AI labs. In the worst case it reports, the provider patched within a week.
FOR THE MEDAL COUNT Medal holders fell from 37 to 19, a 49 per cent drop by our arithmetic, and 15 providers sit in the new Participation Ribbon tier, per the report.
The framework observes: “This entire ecosystem has been shaped by constraints, bottlenecks, chokepoints, and scarcity at every layer.”
As we read it, ClusterMAX points to a constraint inside the compute layer: not only how many GPUs exist, but how many managed clusters meet a basic bar for reliability and security. The framework was published in May 2026 and does not discuss ClusterMAX.
ClusterMAX is one firm’s rating of managed GPU clusters under its own criteria. It does not rate whole clouds, and the sources read include no response from AWS, Microsoft, Crusoe or Together to their tiers.
Neither Patel nor the August article names the provider in the intelligence-agency case or the country. Patel’s talk describes cluster-wide access by an open-source model; the article describes exposure across several providers. Our tier comparisons are against ClusterMAX 2.0; the April 2026 update, 2.1, was not read.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
SemiAnalysis’s ClusterMAX 3.0 gives a medal to 19 of 77 GPU clouds, down from 37 in the previous edition, and 10 of the 19 are Bronze, by our count. AWS, Crusoe and Together are among them, and Azure fell to Silver. Patel said many clouds missed an injected GPU fault until told, and that AI models found published flaws that exposed other customers’ data.
A note on sourcing. Patel quotes are from the auto-generated captions of the Runtime 2026 video, with timestamps; ellipses mark omitted filler words. Ratings and tier history are from SemiAnalysis’s ClusterMAX 2.0 and 3.0 reports and its August security article. Medal counts by tier and all percentages are our count and arithmetic. Security details are kept at a general level.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
