| |
AI ARMS RACE |
THE CLOSE |
Knox Bennett, Lead Strategist, Strategic Compute · September 18, 2026 |
|
|
|
|
| |
| |
|
|
flat · EUPHORIA
V13 C18 Cr17 F12 P11 |
|
|
What moved it: Nothing material sessionless Friday. Credit held. Val+Froth estimated. |
|
|
|
|
|
|
| |
|
|
The Administration Already Took Stakes in Four. Rickards Thinks This Is Next. |
MP Materials, Lithium Americas, Trilogy Metals, USA Rare Earth. Each time the federal government took a direct position, the shares moved. Jim Rickards believes a tiny $2 stock is next on that list. |
|
|
Paid advertisement from Paradigm Press. Not investment advice. Investing involves risk including loss of principal. |
|
|
|
|
TODAY'S BEAT - INFERENCE SCALE |
MLPerf: AMD owns extreme scale, Rubin previews the rack |
512 MI355X vs first peer-reviewed Vera Rubin numbers. |
This morning was SiGe capacity behind NPO and CPO. Tonight is what the same buildout looks like when you put numbers on the scoreboard. |
MLPerf Inference v6.1 is out. AMD submitted a 512-GPU Instinct MI355X cluster - the largest system ever put into the suite - and led aggregate throughput on DeepSeek-R1 and GPT-OSS-120B at that extreme scale. NVIDIA kept the common 8-GPU and 72-GPU racks competitive on Blackwell. And for the first time, Vera Rubin (VR200) showed peer-reviewed preview numbers: a 72-GPU Rubin config nearly doubled a matching GB300 rack on one test path, and a 36-GPU Rubin beat a 72-GPU GB200. |
That is not a single-winner story. It is two scarce slots in one round. Extreme-scale cluster leadership is a different product than next-rack architecture. Procurement teams now have MLCommons-verified Rubin data, not only vendor-hosted demos from the day before. |
The mainstream still wants one ticker to own the inference decade. The second derivative is whether you are buying scale-out cluster efficiency or the next NVL-class rack generation - and whether your software stack can use either. |
| |
| |
512 GPUs is a different question than 72 |
AMD's Crusoe cluster answered the extreme-scale question. NVIDIA's GB300 answers the rack question most buyers actually order. Mixing those scoreboards is how lazy narratives get written. |
|
|
|
|
| |
| |
Rubin's first verified numbers matter |
Preview category or not, peer-reviewed MLPerf is a different artifact from a summit demo on the vendor's own rails. Governance of the benchmark is part of the trade. |
|
|
|
|
| |
| |
ADVERTISEMENT |
|
|
The AI Meltdown Has Begun - Five Steps to Take Now |
Cisco was the Nvidia of the internet build-out and the main supplier of its infrastructure. What happened to it when the dotcom bubble broke is the template here. |
|
|
(Ad) Paradigm Press |
|
|
|
|
| |
| |
THE LONG GAME |
The long game is not crowning a single inference king from one benchmark round. It is matching cluster scale, rack architecture, and software to the workload. AMD showed extreme scale. NVIDIA showed the next rack can jump. Buyers who collapse those into one slide will misprice both. |
Watch how many production RFPs cite MLPerf v6.1 versus summit demos. Watch MI350P and workstation-class cards keep widening the mid-tier. Watch Credit into October while CapEx stays loud. |
|
|
|
|
| |
| |
ADVERTISEMENT |
|
|
Apple, Google and Tesla Were Not Obvious Either |
They are household names now, but the original Magnificent Seven earned that over years rather than arriving with it. Analysts believe the next generation is forming, and seven companies currently fit the same pattern. Free report. |
|
|
(Ad) MarketBeat |
|
|
|
|
|
|
| |
| |
YOU'RE EARLY. KEEP IT THAT WAY. |
The gauge is still 71. The scarce slot tonight is reading two inference scoreboards without collapsing them. Forward this to one person still shopping for a single winner. |
|
|
|
|
The gauge is 71. Credit flat. Valuation and Froth estimated. Written on the morning wake as The Close - thesis-first, not a close rewrite. |
Knox Bennett, Lead Strategist, Strategic Compute |
P.S. If your RFP cites only the largest cluster or only the newest rack, which failure mode are you buying? Reply and tell me. |
|
|
| |
| |
AI ARMS RACE
A Guardian Financial Publishing title |
CONTENT DISCLAIMER
AI Arms Race is for informational and educational purposes only and is not investment advice. Investing involves risk, including the possible loss of principal. |
HOW WE MAKE MONEY
AI Arms Race is free to read. We keep it that way by including paid advertisements and affiliate links in some emails. |
|
|
|
|
|
|
AI Arms Race is a publication of Guardian Financial Publishing (D/B/A) Stable Financial Publishing
Privacy Policy · Terms · Disclaimer |
|
Tidak ada komentar:
Posting Komentar