800 VDC Marks Inflection Point for Data Center Power
Was fun going deep on 800 VDC. I had to talk to my friends in the trade working on these data centers for context lol. Lots of juicy nuggets I uncovered. 800 VDC: The Inflection Point Reshaping Datacenter Power and AI Infrastructure. https://t.co/Ispz18CUmu
Skipping the DTC Playbook, They Launched Soft Basics
They didn’t follow the DTC playbook and get into soft joggers, short shorts, quarter zips, soft polos, etc.. 😂
Understanding 800VDC: Shaping Data Center Power Future
Tomorrow on @DiligenceStack Primer on 800vDC and how to think about the future of power in the data center. https://t.co/HIMjUJ0hIB https://t.co/5Bpqj83kUT
Enterprise AI Adoption Still Under 7%, Not Fully Deployed
Definitely a fad. Fewer than ~7% of enterprises have deployed AI at any kind of scale. 0% fully deployed. https://t.co/DePSverGtr
H100 GPUs Skyrocket in Price, Supply Remains Scarce
Lots of SKUs of H100 commanding much higher prices. Inference has moved to N-2,3 etc compute hardware. And we don’t have nearly enough.
Future Silicon Will Be Built for Agentic AI
From here on out, all silicon architecture designs will be designed with agentic AI in mind. Basically, all silicon becomes agentic workload-focused.
Compute Gap Vast; AMD Poised to Fill It
The compute gap is very large. I'd love to come up with an estimate of how large, will try, but its large. NVIDIA alone can't fill the gap, not sure even TPU scale can help fill the gap at...
MicroStrategy Discusses BE Strategy with Applied Materials
MS on $BE Had a long chat with folks at AMAT on all this last month. https://t.co/KWTiLzZLV5
TSMC's CoWoS Pattern Drives Silicon Photonics Growth
Had a lot of optical discussions at GTC last week. While it may not be one foundry to rule them all, we still think all roads will lead back to TSMC. TSMC COUPE: Why the CoWoS Pattern Is Repeating...
Optical Supply Gap Hits 20‑30% Amid Rapid 800G Demand
Interesting interview with a Sr Engineer in networking at AWS. Optical is ramping fast. Looking to shift to 1.6T but still can't get all the capacity they need at 800G. Diversified across vendors with InnoLight currently the most share....
TurboQuant: When It Hurts vs Helps Memory Demand
Ok, class, listen up. Your homework today is to come up with scenarios where TurboQuant is negative to memory demand and scenarios where it may actually boost memory demand.
Conversational Maps Redefine Everyday Navigation with Gemini
Just used Gemini in Google Maps to find a stop on my way home from SF and it’s amazing. How did we live without conversational maps.
Agentic CPUs Drive Datacenter Growth Amid Specific Workloads
In this report, we detail why, and what workloads specifically, are causing the datacenter CPU growth cycle. A new architecture emerges as agentic native CPUs find their place in the market. Report: Secret Agent CPU https://t.co/jjdrB3k1Bt
Arm Projects $100B AI CPU Market by 2030
$arm believed the dollar TAM for Arm IP, subsystems, and their AGI CPU is ~100b. That’s across customer dollars and agentic CPU dollars down their merchant solution. We do believe the new CPU server TAM across scenarios range from...
Purpose‑Built CPUs Drive New Cloud‑Native and Agent‑Native Markets
Purpose designed CPU for agentic workloads is the reason we will have cloud (web software) native data center CPUs and then agent native CPUs as a greenfield growth category. Good for everyone making CPUs whether Arm or x86. https://t.co/lfWNEf64U5
ARM Expands From IP Licensing to Chips, Boosting Revenue
$arm CEO @renehaas237 emphasizing @arm now sells IP, complete subsystems, and now chips. Tapping into a much larger revenue pool than just licensing should bring $$ upside faster than most have modeled. https://t.co/VKPOyqBzLS
Heavy Computing Demands Lots of CPU Power
Also, since I’m publishing on this topic tomorrow… computer use require A LOT of CPUs.
Excited to See Arm Ecosystem Unveiled at Upcoming Event
I’ll be there. Looking forward to the news and seeing a large part of the Arm ecosystem on display.
TSMC to Depend Heavily on OSATs for Capacity
I do genuinely enjoy some of the company names on Asian exchanges that literally make clear what they do. Global Wafers is another one. That said, TSMC is going to lean heavily on OSATs for capacity support. Interesting times....
Long‑term Supply Contracts Now Standard Through 2028
Yep, and through 2028 now at least. Everyone in the supply chain we talk to is noting how LTA's extend with every customer conversation. Now multi-year is the standard--which is something if you know historical memory supplier agreements.
Elon’s 100k WPM
Still think he was better of JVing with Intel foundry. He will struggle to get ASML machines, even if @substrate is invlolved it is many years before this 100k WPM is remotely possible. I'm sure Elon is taking the...
Testing AI Tennis Ball Machine's Pro-Level Rallies
About to play tennis with this smart AI ball machine. Apparently can simulate pro level rallies. Will get video and report back. https://t.co/YMlVxMRXD8
Price Barrier Leaves Millions Craving Affordable Macs
Our consumer research has always validated a large cohort of people who wanted a Mac but price was too much of a barrier. Most of these respondents were not "heavy" users. Neo success should not surprise. ~1.4B unique customers...
Measuring AI Factories' Value Essential for Agentic Enterprise
As we transition to the agentic enterprise. Stil work to be done to fully model the value of intelligence manufacturing plants (AI Factories). https://t.co/HbYZCQFhhC
On‑prem Token Generators Reshape Edge Hardware Strategy
We think these token generators that can live on-prem and serve tokens to teams are a new factor to consider in modeling where all this hardware goes and how agentic plays at the edge. Good example from @dell here with...
Buy‑side Questions Cycle Durability; Stay Bullish yet Cautious
Talking to a lot of the buy side, it's clear that, despite understanding the constraints and the demand, the central debate is still the durability of the cycle across the board. It is ok to be bullish, but being sensitive...
NVIDIA's AI Factory Expansion Shrinks Competitors' Window
$NVDA post GTC report on The @DiligenceStack The product family is now a full drop in AI factory. NVIDIA Is Expanding the AI Factory and Narrowing the Competitive Window https://t.co/N1Sb2hBqxm
Your Optical View Determines Semiconductor Market Position
Optical was a big discussion in my meetings with many in the semis ecosystem at GTC. This part of the discussion is a debate most underappreciated, as your view on this dictates how it plays out and who is...
Micron
MS note on $MU. I get it, and this both makes sense and makes no sense at the same time. https://t.co/eXaRfWNQ04
Earnings Calls Will Spotlight Memory, Storage Margin Growth
Remember what I said, a few quarters ago, the story will be for memory and storage. Margin expansion. You will hear it all over these earnings calls.
Jensen Highlights Mastery of 2,000‑Provider Supply Chain
Jensen also mentioned that their supply chain consists of roughly 2000 providers. As I wrote in this report, they are masters of the supply chain. https://t.co/9ZCdGgQZVs
Groq Rack Packs 8 LPU, FPGA, CPU, DPU Per
Groq rack is 8 LPU, 1 FPGA, 1 CPU, 1 BF DPU per tray and 32 trays in a rack.
Vera Rack Packs 256 CPUs, 64 DPUs, Liquid‑cooled
Vera CPU tray. 8 CPUs per tray total 256 CPUs per rack. 64 Bluefield DPU (still a CPU) per rack. No wattage disclosed yet for a full Vera CPU rack. Fully liquid cooled. https://t.co/6bbrZEGGuP
Groq’s AI Rack Comb
Groq board. FPGA (likely Altera) and CPU head node (likely Intel) and 8 Groq LPUs on top. Groq rack also on the right paired with VR NVL72. https://t.co/CM4ni9TDyX
CPX Continues Independently, Partnering with Groq to Accelerate AI
Getting some clarification, but CPX is not dead will go forward on its own product path along side Groq. Both have separate, but very specific jobs to accelerate AI.
2027 Market Nears $600B, Networking Over $60B
2027 more likely to be closer to $600B. Networking alone could be north of $60B.
Optical Interconnects Will Dominate Over Copper for Scaling
It does seem copper will be an option in Feynman, BUT, I'd bet the mix shifts heavily in favor of the optical options in this timeline. A comment was made in a session I was in yesterday that once you...
Agentic CPUs Demand Top Single‑Core Performance, New Architecture Needed
As always, solid interview by @benthompson with Jensen. Lots of nuggets, but one key for me is agentic CPUs requiring the best single core performance. Seems likely agentic CPUs will require a different architectural approach. https://t.co/uovoNuv1fg
NVIDIA’s Visual Guide to Scaling Up, Out, Across
In case it helps here is a visual on how NVIDIA describes scale up, out, and across. https://t.co/BVhQbE17Qk
Integrating Groq Keeps In‑House, Not Outsourced
Jensen is articulating what my main concern with Groq was in that as an independent solution why would I outsource all my inference. By integrating it as a part of the solution, it's just disaggregated but not 100% outsourced,...
NVIDIA's Co-Designed System Delivers 35× Efficiency Boost
The new system of compute from NVIDIA. Extreme co-design = 35X more throughput per MW. https://t.co/Z6JtDKckOh
NVIDIA's Massive Installed Base Fuels Ecosystem Strength
Jensen is emphasizing something he has said a few years now on the strength of the NVIDIA ecosystem. Massive installed base and architecture compatibility.
Nvidia Can Tap Every AI Funding Stream
Bear case aside, $NVDA can pull from almost all of these dollar buckets for AI infrastructure. https://t.co/5HKeoRjnJM
NVIDIA's Expanding Stack Inflates Total Infrastructure TCO
One of the more interesting things I'm curious about is how, as NVIDIA expands its infrastructure offerings, all new infrastructure competes for GW. The story has to be about the compounding TCO of the full solution, beyond GPU racks.
OpenAI’s Unified Platform Powers Three Distinct Engines
On the back of the report on Anthropic, publishing our full growth thesis on OpenAI. @DiligenceStack OpenAI: Three Engines, One Platform https://t.co/9inLX4EVvr
Future AI Agents Will Require Vastly More CPU Power
I would also add that agents will need a lot more CPUs.. Via MS note on software in an agentic world. https://t.co/eJpBqngKLu
Apple May Shift to Device-as-a-Service for Education
This got me thinking, this could be how Apple starts to offer devices as a service to things like education, perhaps enterprise someday.
Nvidia Controls 86% of AI Infrastructure Market
In the AI infra semiconductor TAM ~$1.4T estimated for 2030 there are 18 major sub TAM buckets of varying sizes. $NVDA plays in 12 of those 18 and that 12 $NVDA plays in = ~$1.2 T which is 86%...
Vertical Rack Designs Make Optical Interconnects Essential
This is when optical will move to scale. Feynman concept rack from CES, blades stacked vertically, not horizontally. Once rack designs move to this (may be training racks only), then optical becomes a must. https://t.co/CR2fBiX5kH
Few Know I Own Multiple Apple Devices
No. iPad, I think one version of watch if not more. Anyway, not the first time, just very few knew :)