Inference-Era Tokenomics Will Let Compute Print Money
So many of the AI bears were only analyzing ROIC and tokenomics in the training era. Inference era tokenomics will drastically flip this dynamic and those with compute will be printing money. We, and when I say we I mean @mweinbach, is working on something that will model this.
Compute Shortage Drives AMD, Amazon, Cerebras Deal Rush
Chatter being $AMD could get a deal with Anthropic which makes total sense given the compute shortage. I have been saying for a while I think Amazon is a good candidate also for Helios. Bottom line, all available viable compute will...
Intel Nears ATH; Terafab Deal May Redefine Foundry Landscape
Intel ~9 % away from ATH. I also think the terafab deal may be more consequential in many respects than is realized today. Wait until you see my foundry model :)
Margin Expansion Remains Core Strategy for Companies
Holy GMs. I have said this many times but the main storyline for many companies remains — margin expansion.
NVIDIA's Moat: Turning AI Complexity Into Cost Savings
Drastically symplifying, but one take as the strategy/competitive advantage that is becoming more clear from $NVDA is: NVIDIA’s moat is the ability to continuously compress the full AI factory problem into better economics for customers.
From Chip Shortage to Industrial Systems: New Value Drivers
There is a key theme here from Jensen that underscores my recurring thesis I continually unpack @DiligenceStack: The market evolves from a semiconductor scarcity story to an industrial systems story in which energy, networking, packaging, software, and algorithmic efficiency all determine...
UBS Flags Crowded Longs, Calls for More Upside
UBS has $LITE and $CIEN as crowded longs. AKA consensus favorite's requiring more consensus upside/much better results over expectations to move the stock. Sort of feels like many names in optical are getting to be in a similar spot....
TSM Boosts 2026 Outlook to More Bullish
With $TSM tonight, what I wrote in January on 2026 outlook is a good primer. May even be more bullish of a year than we anticiapted even two montsh ago. https://t.co/C0GIvgrEll
ASML Earnings Validate January WFE Outlook
Have a big report coming on WFE as I share a full model on the sector, foundry TAM forecast, and cost per transistor/cost per gb (memory) models. But on the heels of $asml earnings, my base report for WFE...
CRDO: SerDes‑Focused Interconnect Leader, Not Just Copper
Best to understand $crdo as an interconnect company whose vertical orientation, particularly their SerDes, puts them well positioned for every pocket of interconnects. They are not a copper company just now thinking about optical.
Algorithm Glitch Repeats Old Tweets, Stalls Timeline Updates
@nikitabier something up with the algo? Keep seeing a lot of old tweets, and many of the same tweets, vs newer more dynamic information flow. Not as responsive timeline updates either.
Report Shows CPUs Can Achieve 2:1 Scaling
None of this would surprise anyone who read my full report. I even made the case that in what scenarios CPUs could go to 2:1. https://t.co/jjdrB3k1Bt
Apple's Edge Token Lead Makes Windows Price Fight Futile
Will be increasingly hard for Windows OEMs to compete with Apple on price. Not something many had on their bingo card. AND if Mac’s end up being better at edge token generation, any cost argument from IT is...
Dual Copper and Optical Roadmaps Drive Interconnect Success
Guess what is a good business and technology strategy in the interconnect world? Having both a great copper and a great opical roadmap. $CRDO
Compute Constraints Are Compounding; Deeper Layers Reveal New Scarcity
In this moment we are in where constraints in the compute stack are basically “stacking” you have to go one more layer deep to find even more situations of scarcity.
Memory Shortage Pushes Supply Normalization to 2027
Summary of key points from a memory tracker I'm running weekly: Supply normalization now looks pushed to late 2027. New capacity is delayed, DRAM bit supply growth is only ~16% YoY, and that is still below historical norms. Demand Destruction: Memory inflation...
Software Outpaces Hardware: Compute Shortage Limits Progress
Not enough compute is the correct take IMO. Which has quite a lot more implications if you think about it and play that out to its logical conclusion. Software still burdened by hardware’s inability to keep up. Maybe software is...
AI Will Force a Massive, Lengthy Rearchitecture of Everything
All infra, all software, all services, will have to be rearchitected for AI. This is a long cycle.
AI Designs Bee Garden, One‑trip Home Depot
Today's project: Can AI help Ben create his new bee garden structure and only have to go to Home Depot one time. https://t.co/1BpRXN0qBJ
TSMC's Customer Base to Quadruple High‑Revenue Clients by 2027
Doing some customer modeling on TSMC and this stat stands out: In 2023, TSMC had one customer (Apple) above $10 billion in annual revenue. By 2027E, it is projected to have four, including Apple, all above $20 billion.
Semiconductor Market Poised to Reach $2 T by 2030
As insane as it sounds (and again driven by ASP) it may hit $2T by 2030.
Anthropic's AI Layer Poised for $26B Revenue
We articulated this abstraction layer Anthropic is going after and how this turns into $$ in our full report on Anthropic and their revenue potential. The financial model we built end of February had them at $26B in 26...
Transistor and Memory Costs Are Rising Astronomically
When I say the cost to make transistors ($ per billion) has gone up.. I mean its gone UP.... Memory is even more insane, cost per gb is 🤯
Anthropic Positions Claude as Universal Productivity OS
Anthropic is on a full-fledged pursuit of usurping the productivity value from all other applications/software and taking it for themselves in Claude as the productivity OS.
Memory Insights and Skippable Tim
Whole video pod of @T_h_e_Circuit and the points I made earlier in the week on memory, we discuss in the chapter timelines you can skip to.
Apple Boosts TSMC SoIC Capacity for AI Servers
Morgan Stanly on Apple custom ASIC for cloud/hybrid compute (chip code-named Balta): "Apple is materially ramping SoIC capacity at TSMC, pointing to a major push in Apple silicon for AI servers. TSMC (covered by Charlie Chan) is expanding its SoIC (System...
Mythos May Not Be Exclusively Trainium
We suspected this, which was why I hadn't yet commented on the sentiment that Mythos is a blackwell trained model. May not be 100% trained exclusively on any GPU/XPU in actuality.
Liquid Cooling Unlocks AI-Scale Compute Capacity
Part 2 of the future of the datacenter series. 800 vDC + liquid cooling becomes a co-design situation with compute infra decisions. The key is the compute capacity it unlocks. Liquid Cooling: The Thermal Prerequisite for AI Infrastructure...
AI Slashes Amgen's Drug Development Timelines up to 70%
Morgan Stanley met with Amgen (biotech) management on how AI is benefiting their workflows. Key findings: - 8%: Approximate industry drug-development success rate Amgen cited as the baseline AI could help improve. - 10–14 years: Typical drug-development cycle time Amgen cited...
Rising Memory Costs Threaten AI ROI and Market Confidence
While we talk about memory prices 📈, and while that is true (great for memory names), I am worried about what this does to capex costs and in return on AI $$. If the cost to serve AI goes...
AWS Poised for Its Most Consequential Growth Phase
Seems a good time to resurface this report on the back of the Amazon shareholder letter. The AWS Acceleration Thesis: Why Amazon's Cloud Business May Be Entering Its Most Consequential Growth Phase https://t.co/nNOw2zxMk9
Intel Missed Chance by Denying Gelsinger's Foundry Investment
Can you imagine if Intel's board let @PGelsinger invest in the foundry capacity he asked them for? I agree, it seemed crazy at the time, but hindsight is 20/20.
Points 5 and 6 Define Cycle Durability, Not
I liked the $AMZN shareholder letter. I take point 4 with a grain of salt (trn is largely AWS as a customer and Anthropic). Points 5 and 6 are most important for the "durability" of the cycle debate. https://t.co/X7zQDnJ6YM
TSMC and Top Memory Makers Set Prices, Customers Agree
TSMC, and the three memory names right now, are the primary beneficiaries of a name your price market dynamic. Meaning they can literally name their price and customers will pay it.
Barroom Chats Reveal Data Center Construction Insights
This is how you do it. One of my other favorite anecdotes was a friendly on the buy side who sent analysts to research data center builds (delays, issues, etc.,) and they got info talking up contractors at the local...
RMBS Spotlight: Labs’ IP Accelerates Memory Industry
I heard people may be starting to look into $RMBS. I published (in February) an interview with their SR Fellow of Labs on how the memory industry is being pushed forward with their IP. Hope it helps. https://t.co/nrz9hxwGHR
Researching for DiligenceStack Pushes Me Into New Tech
I like how writing @DiligenceStack forces me to go do research I wouldn't normally have to do on the client side. But it's making me go from from the weeds of semis and now to the soil. All of...
2026 Market Forecast Jumps $200B, Extending ROIC Horizons
It is wild that about two months ago, consensus memory in dollars for 2026 was ~$600B and now people are saying north of $800B... Very challenging to square this with ROIC timelines now, as they will have to lengthen, and...
From $1T in 70 Years to $2T in Five
What no one saw coming about this, since most $$ estimates were on volume not ASP, is just that - this is simply costs going up. It took >70 years to hit $1T and it will go from...
TSMC Poised to Dominate Silicon Photonics Like AI Packaging
I'm glad TSMC COUPE is getting more attention of those in the weeds in optics. Our view: "TSMC is increasingly positioned to occupy the same role in silicon photonics that it came to occupy in advanced AI packaging." https://t.co/WMdxP2tF3R
Transformers: Powertrain Driving Grid Upgrade Surge
Why we who track the AI buildout should care about transformers. They are the “powertrain” of the grid upgrade cycle, with demand being driven by three forces at once: aging grid infrastructure, the shift toward renewables, and a sharp increase...
Memory Dominance Determines AI Compute Winners for Years
If the winners and losers in AI compute come down to who gets the most memory in the most ridiculous memory-constrained market we have ever seen.... For the next 3-4 years.... Predicting those winners is pretty easy.
Premium Pricing Ensures Big Spenders Dominate AI Landscape
One wonders if premium models/token pricing will always exist. If so, then those companies who spend the most will have distinct advantages.
Memory and Storage Converge Into a Single Layer
New report on @DiligenceStack. Companion to the memory deep dive as co-optomization is a memory and storage story. Storage Wars: When Memory and Storage Collapse Into One Layer https://t.co/ozDZpHYNi3
JPM Predicts AVGO Valuation Exceeds $120B Post‑earnings
After $AVGO earnings, I said ~120B was the likely number for 27. Now, looks like higher is likely. This from JPM. https://t.co/OvxZg1VUWr
Terfab Ranks Among Top Three WFE Spenders for 1GW Compute
Putting Terfab into perspective. Based on stated goals: 1GW of compute requires about 50kwpm of logic, 140kwpm of DRAM, ~30kwpm of NAND, and ~35 kwpm of CoWoS (advanced packaging). That translates to about $30bn of WFE per 1GW produced...
Anthropic Poised to Outpace Salesforce in Revenue
Seems probable that Anthropic passes Salesforce in revenue this year. I know not technically SaaS, more like intelligence as a service, but the fastest scaling enterprise software company in history is telling.
New Results Closely Match My Anthropic Model
They are tracking very closely to the model I built on Anthropic in this note. https://t.co/cspnwjWdqN
Apple Poised for Record Mac Sales with Neo, Mini
Between MacBook Neo and Mac minis (thanks to Open Claw) Apple may have a record year of Mac sales.
New Model Updates Cost‑Per‑Transistor Economics for AI Era
Working on building a new cost per transistor model with updated economics since it's been a few years since I have seen one. So many more dynamics are needed than the ones from a few years ago, pre AI...