Silicon Valley’s Starving for Compute

Last month, I had conversations with people up and down the Bay Area as part of my trip to Hot Chips 2026.

Then I boarded a flight to Taiwan for Semicon Taiwan, and I had some more conversations there, though not as much as I was getting pretty tired. It's been

a year since my last trip to the United States. A lot has happened. Year four in

States. A lot has happened. Year four in the AI boom, and a string of recent developments has given the boom substantial new energy.

the shift to long-running AI agents and agent swarms. The pursuit of RSI, which stands for recursive self-improvement and does not refer to a risk condition,

cyber security and the hugging face hack, which I won't talk about here, and the epic fight for compute. Another

year, another vibes video about things in the Silicon Valley and Taiwan. This

video is brought to you by the Asianry Patreon. Before we begin, I want to say

Patreon. Before we begin, I want to say that I'll be returning to San Francisco for Semicon West 2026 to give a short talk at the Test Vision Symposium on

October 15th, 2026 at about 12:45 p.m.

I'm still working on that talk, but I have a pretty good feeling that it is going to be about test. After all, it's at the test vision symposium. If you

have the time, please come to watch it and say hello afterwards. I would love to chat with you guys. Now, on with the show. The visit started with the Hot

show. The visit started with the Hot Chips show, of course. Hot Chips is bigger than ever before. This is my third year visiting. My first year there, I remember getting off the plane

and heading immediately to Stanford. The

auditorium was super warm. I was

jetlagged and so I ended up falling asleep, but I managed to get away with it in part because the back of the auditorium was so empty, perhaps 60%

full at most. Flash forward to today and falling asleep again like that would not be possible. The whole place was

be possible. The whole place was absolutely packed and to get an open seat you have to step over many other individuals way more finance people this

year too. One major theme I gathered

year too. One major theme I gathered from people at the show is the forthcoming invasion of AI into the EDA space. I did a previous video about AI

space. I did a previous video about AI enhanced EDAs and I recommend you jog over to that one if you want that.

Another major theme was silicon photonix. Last year, hot chips lined up

photonix. Last year, hot chips lined up four photonix presentations back to back to back to back and it was fire. While

it was not pre prominently featured this year, there was still many Photonix people at the show. My favorite this year was when Micron, Samsung, and SKH

Highix went up to discuss memory. The

memory sessions started at Sunday morning 9:00 a.m. yet were absolutely packed. 500% or whatever stock booms

packed. 500% or whatever stock booms will do that for you. Last year for tonics, this year we have the memory folks. I wonder what cluster we're going

folks. I wonder what cluster we're going to see next year. That's alpha, I reckon.

The main event, the one everyone seemed to be waiting for, was Open AI presenting its new AI chip, Jalapeno.

That presentation was short and sweet and served up plenty of interesting and spicy food for thought. Last year, I wrote that the window for new AI chips has closed with the exception of

Google's TPU. I take that back now. I

Google's TPU. I take that back now. I

was wrong. AI Infest chips are so hot right now. Open AI brought some heat to

right now. Open AI brought some heat to the show, but outside there lurked a cotari of new AI chip startups including Etched, Matt X, Posetron, and Fractile.

Etched in particular made a splash at the show despite not having a presentation. Jeff Smith of Edge seemed

presentation. Jeff Smith of Edge seemed to make it his mission to crash every Q&A. Every time asking a smart yet

Q&A. Every time asking a smart yet somehow dissatisfying question. They

also handed out a grab bag with a nice hat in it. I will wear this hat. Thank

you. These companies are hardware companies. They're not just designing

companies. They're not just designing the chips inside computers. They're also

putting together the racks and systems running these chips. Meaning that they have to roll up their sleeves and get messy in the lab. This is not a trivial task. I hear often about the challenges

task. I hear often about the challenges of acquiring top talents and I am impressed with the execution.

So how many of these companies survive?

Right now there's a compute shortage.

I'll get to this later. So I fall back on the disputed words of Nathan Bedford Forest. Get there first with the most

Forest. Get there first with the most men. It all depends on execution and

men. It all depends on execution and what can be achieved in a constrained supply environment. Like can you get

supply environment. Like can you get your hands on HBM? Should you be getting HBM in the first place? Can you tape out without major errors? If not, then what

can you do about it?

Let me jump around a bit and fast forward to Taipe's Nang District as I attend Semicon Taiwan. I've been

attending this show since 2022. And back

then, Semicon Taiwan 2022 was a humble affair. Chill. Though that might have

affair. Chill. Though that might have had something to do with the CO restrictions still in place back then.

The welcoming booth was a simple green lit setup. The show floor was relaxed

lit setup. The show floor was relaxed with a few engineers or salespeople lingering about and they ran the information sessions in Mandarin Chinese. Things have changed. This year

Chinese. Things have changed. This year

was mayhem. Every room in both Tynx 1 and Tinx 2 was filled to the brim. The

hype and showsmanship was to the max.

The show had traditional showg girls at the booths, but also these cute attendees walking around in bunny suits and fake wafers. They had this dream fab

concept where people can dress up in the bunny suit and then go into the simulated fab and see tools there. You

have to sign up for the chance to go and it was absolutely filled up. No chance

for me to take a look. First timers to Semicon Taiwan might be surprised to discover that TSMC or ASML do not have a booth at the show. I'm not sure why ASML

is not there. I do see their job recruiting ads in Taiwan, but TSMC dominates the show despite not having a booth. Their teams are walking the floor

booth. Their teams are walking the floor going up to new vendors and investigating new technologies. And many

of the info sessions are headlined by a TSMC executive who goes up in front of the group and broadly lays out their problems. And then it hits you. Semicon

Taiwan is TSMC's request for proposal to the whole semiconductor industry.

The technology line is clear. The future

of semiconductor manufacturing is to go bigger. Right now, AI chips are

bigger. Right now, AI chips are straining the industry's existing limits. Most notably, the retical limits

limits. Most notably, the retical limits that restrain how large of a dive that an existing lithography machine can print. Many emerging technologies are

print. Many emerging technologies are either about breaking past those limits or scaling the rest of the ecosystem to accommodate. That means advanced

accommodate. That means advanced packaging. TSMC has a slide showing them

packaging. TSMC has a slide showing them preparing battleship size chips spanning up to 14 reticles. The industry is responding with typical rapid speed. For

instance, two years ago, I did a video about panel level packaging, a theme of the 2024 version of Semicon Taiwan.

Interposers are for hosting all those chiplets and memories, and they're usually made from silicon. But AI

systems are so chunk now that the industry wants to use panel interposers that are square shaped and produced from glass.

A couple years ago, panels were all talk and slideshows. Can TSMC produce these

and slideshows. Can TSMC produce these panels without warping and cracks? I'll

believe it when I see the chips, but the ecosystem is getting ready on the floor.

Their handler robots and other items accommodating square panel sizes. I

reckon that this chipmaxing trend will turn out to be one of the most far-ranging technology transitions in the semiconductor industry's history.

Basically, since the adoption of 300 mm wafers, I am all for it.

Back to the bay. I was last in the Bay Area in September 2025, which was an entirely different time. And since then, long-running Agentic AI has taken hold

across Silicon Valley, triggered in part by Anthropic releasing more capable models like Opus 4.5 in the November 2025 time period, as well as the Open

Claw thing. In February 2026, prices

Claw thing. In February 2026, prices start to rise as players scramble to lock down compute, led by Anthropic, of course, which signs big deals with

SpaceX, Google, and so on. At the time of my trip, I'm told that the spot market has completely vanished. For

Neoclouds, this is the greatest thing since sliced bread. The phrase Neocloud was coined by semi-analysis Dan Nishbal in April 2024 to replace what had been a

more profane term. It describes a pure play cloud provider that resells AI training and inference compute. Examples

include Coreweave and Nebus. These

Neocloud firms believe that the comput shortage will not let up for at least one to two years. In this tenant, they have complete certainty. I've met

several people unrelated to one another who are trying to start their own Neocloud.

For the AI startups trying to train models, this compute shortage represents a business risk. Imagine how pleased investors would be to hear that no

product is coming until midway through 2027 because there's not enough compute.

For the VCs invested in those startups, the new way of helping out is tracking down compute. I've heard stories of guys

down compute. I've heard stories of guys chasing down a thousand GPUs here and there, scrging up compute in the most random of places like East Europe.

It is interesting to compare that with the situation inside Anthropic and OpenAI. They are feasting. Dylan

OpenAI. They are feasting. Dylan

recently said on the latest Dashpod that the two AI giants have about 5 GW of compute. Considering the wealth of

compute. Considering the wealth of compute, one might guess that a good amount of it is being wasted inside these companies. Anecdotal accounts do

these companies. Anecdotal accounts do seem to confirm that this embarrassment of riches is not particularly being used with the utmost efficiency.

One question I liked to pose to the AI people I spoke to was will Google return to the frontier. The broad agreement was no though the explanations are a bit squishy. A few pointed out that Google

squishy. A few pointed out that Google deciding to sell off a lot of their compute to anthropic is basically saying that they do not believe it either. It's

a big turnaround from when Gemini 3 got some people believing that open AI was cooked. In the end, it feels like and I

cooked. In the end, it feels like and I was told as such that Google suffered from a substantial talent loss as well as a culture that was not conducive to frontier AI.

What that means I shall leave for some future technology historian with more perspective and Harvard degrees than I do. So if Google is not making it back

do. So if Google is not making it back to the frontier, then who in the United States, let us set aside mainland China, is most likely to join anthropic and

open AI at the frontier? The consensus

seems to be SpaceX hacks AI. Elon does

not strike me as a man who likes losing and he seems sufficiently motivated.

After all, he spent more money buying cursor than he ever did on building rocket technology. And to get talent, he

rocket technology. And to get talent, he was pulling all the same recruiting techniques that Meta did. And the Cursor guys seem pretty sharp, too. They're

undoubtedly the ones behind the recent Grockbot product that I'm told is pretty impressive, though note I've not used it myself.

One thing that especially fascinates me are these things called Neolabs. These

are basically pre-revenue research for startups funded to explore an intriguing AI idea. They're often defined by elite

AI idea. They're often defined by elite researcher founders and little if any business plan other than to discover a brand new structure. There are a number of these out there. The first and

perhaps most famous is safe super intelligence which raised a billion dollars but has not yet released a product or model. The term is a vague one. They include recursive

one. They include recursive intelligence, not to be confused with another company named recursive intelligence but with an I, humans and

core automation and flapping airplanes.

Some other AI labs with a product or two out include Sakana AI in Japan, thinking machines or reflection AI. There are a few more domain specific labs in areas

like material science, world models or robotics like physical intelligence, periodic labs or generalist AI. The

knee-jerk first take is that this whole asset class is dumb. Why are you giving these people all this money? No revenue,

no immediate concrete plan to make revenue. But I can kind of see it.

revenue. But I can kind of see it.

Historically, to produce groundbreaking science and usher it to the commercial realm, you look to research universities, government laboratories, or something like a Bell Labs. Bell

Labs, low-key funding R&D for everyone by licensing out their patents for decades. But many universities either

decades. But many universities either have had their funding cut or their talents rated by industry. AI is

probably the first major technology of the last 50 years to have emerged without any government involvement. And

in this AI scaling era, research is closed and kept secret. So why not get Nvidia who has the compute and the cash and a bunch of VCs to fund some

asymmetric bets? OpenAI Anthropic are

asymmetric bets? OpenAI Anthropic are unlikely to spend money and research on these ideas. They're sort of locked into

these ideas. They're sort of locked into their existing transformercentric paradigm. So give some super smart

paradigm. So give some super smart people money and see what they stir up.

Failure is likely, but most of the funding is in the form of compute anyway. And if someone does hit on a

anyway. And if someone does hit on a gold mine so compelling to force the giants to switch over, like more data efficient architectures, then the giants will probably buy out the discoverer at

some eye watering price that pays it all back. In my opinion, it's better use of

back. In my opinion, it's better use of Nvidia's cash flow than just more share buybacks. Apple, I'm looking at you

buybacks. Apple, I'm looking at you here. And to be honest, I low-key love

here. And to be honest, I low-key love these companies. It's what makes Silicon

these companies. It's what makes Silicon Valley awesome. a gang of ultra smart

Valley awesome. a gang of ultra smart people working with the force of Lenin to create something new. And I don't mean new as in another SAS. If you work for one, email me and I'd love to come

by and chat during my next visit. One

major question that I wanted to examine yet again while in the United States was customers are burning tokens for sure, but are they getting the ROI? After

speaking to so many people across the ecosystem, the answer remains nebulous.

We don't know. Nobody has measured it.

We're all running on anecdotes. I do

think that the lowering of barriers to code will impact knowledge work. Back in

the day, financial firms hired lots of reporting analysts to help management stay a breast of their positions. Today,

that can all be done with two young smart analysts and claude coding up a dashboard. This I have seen of the AI's

dashboard. This I have seen of the AI's other capabilities like making spreadsheets or presentations, items that seem to be frequently advertised in the marketing. I haven't seen it make

the marketing. I haven't seen it make noticeable impact. Perhaps it just takes

noticeable impact. Perhaps it just takes more time. Perhaps it's an intelligent

more time. Perhaps it's an intelligent thing. OpenAI recently published a blog

thing. OpenAI recently published a blog post discussing how agents have accelerated research inside the company, but they still peg the AI's capabilities

as like an intern. It executes daily tasks. We remain far from an actual AI

tasks. We remain far from an actual AI researcher. I'm told that AI still lacks

researcher. I'm told that AI still lacks a human's research taste, though they believe that it perhaps can be achieved with enough scale. We shall see. One

thing that certainly seems to be the case for everyone in SF is how hard they work. If models are really so competent,

work. If models are really so competent, then why does everyone say that they have more work than ever? Maybe rather

than automating away work, AI just gives us more work to do.

The thing that gets me is how everyone is saying compute shortage, shortage, shortage and for the next 2 years. And

it makes me wonder if maybe things might not turn out the way people expected. My

math is not super sophisticated. I just

think that there's too much supply hitting in 2027.

It is estimated that at the end of 2026, usable AI data center capacity will be about 15 gawatt. Open AAI and Anthropic have about 5 GW each of that. there is a

shortage and the supply chain responds.

That 15 gawatts at the end of 2026 is estimated to surge to about 45 to 55 gawatt end of 2027. So about 30 to 40

gawatt of incremental capacity. Open AAI

and Enthropic will share about 30 gawatt of that between the two of them. Assume

half of that goes to training. So 15

gigawatts of compute for inference together for the big two. Right now,

each gigawatt of inference is modeled to bring to bring in about $50 billion of revenue. This number is from Dylan Patel

revenue. This number is from Dylan Patel on the Dashesh podcast, but we can also project it from Anthropic's $1.25 billion per month deal for SpaceX's 300 megawatt. And Dylan believes that

megawatt. And Dylan believes that smarter models could eventually push that to 100 billion per gawatt. Use that

number between the two giants. 2027

should produce 750 billion to$ 1.5 trillion of revenue for both companies.

Brother, that is a lot. Most people

agree that the two are going to do about 150 to 200 billion this year, though it might be lower. Anthropics S1 coming out literally any second this or next week

can help add some color to that. Anyway,

I can kind of see 400 to 500 billion for the two next year. But 750 billion to 1.5 trillion is two Walmarts at the higher end and two Toyotas at the lower end. It strains credibility and I just

end. It strains credibility and I just don't think they get there. I know many might disagree with me. Even if the models start curing cancer and discovering superconductors tomorrow, I am betting that the revenue number is

just wrong and we get far lower than 100 or even $50 billion per gigawatt, probably a fraction of that. The issue

is that the infrastructure investors need to make money back on their data center investments.

How about we look at it from the infrastructure side? Let us say that 40

infrastructure side? Let us say that 40 gawatt total hit in 2027. Current rental

prices sit about $20 million per megawatt, which correlates to a 3-year payback. Anyway, this means that those

payback. Anyway, this means that those 40 gawatts need $800 billion paid to the infrastructure providers internal and external. That is just 2027. I don't

external. That is just 2027. I don't

think they get there either, which means rental prices must fall and someone somewhere probably takes the bath. These

math exercises probably can be boiled down to the following. The supply chain is going to bring on a lot of compute in 2027. Most of it in the second half of

2027. Most of it in the second half of the year, and I bet it tips us from a shortage to a glut. I'm not saying that the AI bubble's popping, just that we

will soon be a wash in compute. We're in

2026 a gentic demand era, but with late 2025 supply, supply will catch up and probably overshoot.

But then maybe the next phase of AI growth ignites and demand swells again.

Persistent AI perhaps. I want to thank everyone who took their time to sit down with me and giving me a window into their world. If you recall my last Hot

their world. If you recall my last Hot Chips Vibes video, the one about Silicon Valley doing hard things again, I mentioned that someone predicted that AI will solve one of the Millennium

problems over the next year. Now, as I write this, there is news that an advanced internal model at OpenAI has produced a proof for the Navier Stokes existence and smoothness problem. My

friend got it right on the dot.

Unfortunately, what should have been a remarkable achievement has been marred by accusations of plagiarism and data theft. This grounds well of negativity

theft. This grounds well of negativity is indicative of another thing that I'm starting to see more of on the internet and in ordinary people. Anti- AI

sentiment. A good number of people hate AI either because they hate the data centers or because it represents everything they hate about the billionaires or because they think it is stealing real quote unquote human

knowledge or because they don't think we need it in our society.

Regardless of why they feel it, the sentiment is growing. And I have a bad feeling that this issue can win elections. SF operates in a bubble. I

elections. SF operates in a bubble. I

get it. The city literally swirls inside its own microclimate.

But the people living inside should pay closer attention to those feelings growing outside of the SF bubble. Things

can metastasize with scary speed. I

suppose if people are afraid of AI extinction, then this is the best way out, right? All right, everyone. That's

out, right? All right, everyone. That's

it for tonight. Thanks for watching.

Subscribe to the channel. Sign up for the Patreon. And I'll see you guys next

the Patreon. And I'll see you guys next time.