Dylan Patel thinks vertical integration does not yet win hyperscaler orders in China, but could matter more as power gets scarcer and power architectures change.

It is too early to say whether vertical integration will become a meaningful advantage in China’s datacenter market. For now, it gives integrated players more control over cost and schedule, but hyperscalers still choose vendors mainly on resources and track record. That could change as power becomes harder to secure and the power system inside the datacenter is redesigned.

Dylan Patel argues energy access is becoming part of what datacenter developers sell, and power architecture is shifting to high-voltage DC with batteries replacing UPS.

Across different markets, access to energy is becoming part of what a datacenter developer sells. Second, the power architecture from the grid connection to the rack is being redesigned. In the US, the industry is moving toward higher-voltage DC power, while battery backup units (BBUs) and BESS are starting to take over the role of UPS.

Dylan Patel notes China's datacenter market is fiercely competitive, with rents halved, capex down 25%, and Inner Mongolia power among the world's cheapest.

China’s datacenter market is highly competitive. Rents have roughly halved over the past five years, to about $40/kW/month excluding power. Capex has fallen about 25% over the same period, to less than $3mm per MW. Power in Inner Mongolia costs about $0.055/kWh all-in, among the lowest in the world.

Dylan Patel says Google's TPUv7 now matches or beats NVIDIA on perf per dollar in agentic inference, which has pushed NVIDIA into an urgent engineering response.

TPUs delivering better performance per dollar or competitive performance per dollar is a new experience that NVIDIA senior leadership is unwilling to accept. As soon as TPU posted its first AgentX result on Qwen3.5 397B FP8 showing that even TPUv7 FP8 could beat GB300 NVL72 FP4, NVIDIA tasked a strike force of human engineers and agent swarms to work 997, late into the night until 2am local time and on weekends, across USA and mainland China.

Dylan Patel warns that bureaucracy at NVIDIA is slowing its response to the open-source community and is among the biggest risks to its software moat.

There needs to be an cultural shift in massively reducing bureaucracy at NVIDIA. This is one of the biggest risks to NVIDIA Moat that Jensen has to tackle. In our experience and many customers experience, Nvidia is now slower moving than AMD, Google, and Amazon to external suggestion, git commits, and feedback.

Dylan Patel notes that without native FP4, TPUv7 still trails FP4 GPUs on tokens per dollar until TPUv8 arrives in Q1 2027.

FP4 GPUs still deliver more tokens per dollar at most speeds, as TPUv7 has no native FP4 support, and we will need to wait until Q1 2027 for TPUv8 with FP4 support, as tracked in our Accelerator & HBM Model. At 100 tok/s/user, GB300 and GB200 FP4 deliver 3.6x and 3.5x the tokens per dollar of TPUv7.

Dylan Patel says labs with large TPU fleets co-design model attention shapes for TPUs, so a GPU-optimized model like Muse Spark Open understates TPU potential.

Note that AI labs with lots of TPU capacity, such as Anthropic and Google Deepmind and soon OpenAI, typically co-design their models such that the attention shapes would fit nicely on TPUs to avoid this issue.

Dylan Patel predicts NVIDIA may regain a slight lead at FP4 but cannot retake state-of-the-art at FP8 for some models versus TPUv7.

We do think that with enough agent swarms, NVIDIA could regain a slight performance lead on GB300 NVL72 FP4 vs TPUv7 FP8, but on NVIDIA FP8 versus TPUv7 FP8, we believe that it is not possible for NVIDIA to regain SOTA for certain models.

Dylan Patel says SemiAnalysis expects $11 trillion of global AI capex by the end of 2029, with about $7 trillion of it financed by debt, raising the question of who pays.

The other question is sort of who pays for it. So one of the things that we’ve said is by end of 2029 we expect $11 trillion of capex globally on AI.

Dylan Patel argues SpaceX may now be the world's fastest datacenter builder, outpacing even Chinese peers, citing Colossus 2's near-2GW addition in under a year.

We now believe that SpaceX may be the fastest datacenter builder in the world, even more so than Chinese peers. Watch how Colossus 2 adds nearly two Stargate Abilenes (2GW) in just under one year!

Patel says SpaceX's ~13GW 2027 target is unprecedented in scale but feasible, with over 9GW of capacity already identified.

SpaceX has even revised its estimate up to ~13GW in 2027. This would be the largest amount of datacenter capacity ever built by any one company in that timeframe, roughly equivalent to all of Google’s datacenter capacity from 1999-2026 in just 1.5 years. We think it’s possible and have already identified >9GW of capacity.

Patel notes SpaceX has halved typical greenfield datacenter build times to about four to five months, a record at scale.

To our knowledge, this would be the fastest greenfield datacenter ever built at scale. While most are fighting for 8-12-month build times, SpaceX has cut them in half.

Patel attributes SpaceX's speed to elite hardware engineering, lean labor of ~3k workers per GW, and Tesla-made modular MEP equipment.

One thing often overlooked is that the Elon companies are some of the best hardware companies in the world. We mentioned in our previous newsletter that their labor can run at 3k workers per GW thanks to their top-tier engineers. However, we also believe that Tesla is manufacturing modular MEP equipment, allowing SpaceX to deploy so rapidly.

Patel believes Colossus 2 will pair a ~6GWh BESS with gas turbines to cover a roughly 300MW power shortfall after turbine restrictions.

With the BESS nearing 6GWh, we think the site will orchestrate gas and BESS to increase baseload, while the turbines cycle through charging them. If the datacenters are designed right (and our IT MW assumptions are correct), this could work.

Patel has medium-high confidence that SpaceX's TeraFab chip fab project will also include a datacenter, possibly for third-party rental.

Lastly, we have medium-high confidence that TeraFab will have a datacenter too. That’s right, the 10GW, 100M sq.ft chip fab venture from SpaceX will likely have a datacenter.

Jensen Huang said Nvidia's revenue per gigawatt rises from about $18B for Hopper to $25B for Grace Blackwell and $40B for Vera Rubin, with productivity improving several-fold each generation.

In the case of Hopper, we were at about $18 billion per gigawatt. For Grace Blackwell, we're about $25 billion per gigawatt. And for Vera Rubin, it's about $40 billion per gigawatt. And the productivity is X factors increase in each generation.

Jensen Huang said customers are racing to adopt the next generation of Nvidia systems as quickly as possible.

And so customers want to race to the next generation as fast as they can.

Jensen Huang says AI agents are transforming how every company works, and that capability and security must advance together.

“AI agents are transforming how every company works, and capability and security must advance together,” said Jensen Huang, founder and CEO of NVIDIA.

Musk said Tesla will wind down Cybertruck production at Gigafactory Texas around mid-2027 to free capacity for more Cybercab production.

Tesla will be winding down production of the Cybertruck at Gigafactory Texas around the middle of next year to make room for increased Cybercab production,

Musk quipped that space-based data centers look appealing after the midterms, amid worries about a regulatory slowdown on data center construction.

"Data centers in space looking pretty appealing right now," he added in another post.

Musk quipped on X that space-based data centers look appealing after the midterm backlash against terrestrial data centers.

"Data centers in space looking pretty appealing right now," Musk wrote on X on Wednesday.

Musk posted on X that space-based data centers look appealing after Texas election results raised the prospect of data center restrictions.

“Data centers in space looking pretty appealing right now,” Musk wrote, referencing his ultimate ambitions.

Musk suggested on X that space-based data centers look attractive after Texas Democrats won and pledged to freeze data center approvals.

"Data centers in space looking pretty appealing right now," Musk wrote on X after the results.

After the Democratic midterm sweep, Musk quipped on X that space-based data centers look appealing, hinting at escaping state-level AI data center restrictions.

Data centers in space looking pretty appealing right now.

Amodei says he is willing to be mocked for Claude's strict biology safeguards rather than risk Claude being used to cause mass casualties.

I would rather they make fun of me than wake up one day and discover that someone used Claude to kill a bunch of people.

Amon says High Bandwidth Compute is beating Investor Day targets, with about 0.3 pJ/bit versus a 0.5 target and yields above expectations, and is especially efficient for decode.

We're very happy with the progress we're making on HBC. We're actually doing better than what we said on Investor Day. We had a target on the design to about 0.5 picojoules per bit. We're doing better in the lab. We're doing about 0.3. Yield is exceeding our expectations.

Amon argues the smartphone is the essential device for agentic AI because it holds user context, preferences and permissions, and will drive more compute content per device.

The smartphone is an essential device for agentic experiences, as it's the device that knows people best. It's always with them and has their context, preferences, and permissions. The device is also evolving to simultaneously support what the user is doing and what agents are doing on the user's behalf.

Amon says many uncertainties—licensing renewals, loss of Apple, and memory pricing hitting the handset market—are now behind Qualcomm, with handset units expected flat to down in 2027.

We closed fiscal 2026 with many of the uncertainties that have been weighing on our business now behind us. Specifically, QTL license renewals, the loss of Apple as a customer, and the negative impact of memory pricing on the size of the mobile industry.

Amon welcomes the jury finding that Arm breached its license agreement and acted anticompetitively, calling it a victory for innovation and competition.

We're gratified that the jury found Arm breached its long-standing architecture license agreement, ALA, with Qualcomm, failed to honor its commitments, intentionally acted to disrupt Qualcomm's business, and engaged in unfair competition both as a supplier and a competitor.

Amon says the Starlink 5G win reflects an opening to innovate in wireless infrastructure beyond dedicated equipment, via Open RAN and AI-native approaches.

The industry has been running on dedicated equipment. I think we have seen a lot of trends of technology ranging from Open RAN to AI native as an opportunity to innovate in this space.

Amon said Qualcomm's data center constraints come from supply rather than demand, and data center revenue is expected to reach $5 billion in fiscal 2027.

CEO Cristiano Amon said data center constraints stem from supply, not demand, with data center revenue seen reaching $5 billion in fiscal 2027.

Amon says Qualcomm's $5 billion data center commitment is constrained by supply rather than demand, and he would invest more if he had more resources.

And I'll tell you right now, regardless of all this noise, I think our $5 billion commitment is not demand-driven. It depends entirely on the offer. If I had more resources, I would probably make more than $5 billion .

Amon argues Qualcomm's low-power computing DNA positions it to help data centers cut total cost of ownership.

What we do in data centers, if you think about a company like Qualcomm, our DNA is low-power computing. In data centers, there will be a need to reduce the total cost of ownership (TCO).

Amon says Qualcomm's $15 billion FY29 data center target rests on new inference-focused technologies with minimal compute density and power use.

when we think about our $15 billion forecast for FY29, there are a lot of new technologies coming out, and we're focusing on inference, decoding, and trying to do that with the lowest possible compute density and power consumption.

Amon said Qualcomm's handset weakness stems from supply constraints, not weak demand, and that more supply would have improved results.

"We don't have a demand issue; it's a supply issue," CEO Cristiano Amon said. "If I had more supply, I would probably be doing better."

Amon said fiscal 2027 begins Qualcomm's non-handset revenue inflection, including the data center ramp.

“Fiscal 2027 marks the start of our non-handset revenue inflection, including the ramp of our Data Center business,”

Amon reaffirmed that Qualcomm's share of Samsung's next-generation Galaxy flagship chips is now north of 70 percent.

“Absolutely,” he said. “The baseline was 50 percent. This baseline is no longer 50 percent. It’s north of 70 percent.”

Amon says data center is a top strategic priority for Qualcomm, built on long-term customer relationships, a broad portfolio and open software.

Data center is another area where we made considerable progress this quarter. it is 1 of our top strategic priorities and a major focus of my own time, as we build this business around long term customer relationships a broad product portfolio in an open software ecosystem.

Amon says Qualcomm's high bandwidth compute technology beats its power-efficiency target in the lab and is especially efficient for decode workloads.

We had a target on the design to about 0.5 picojoules per bit. We are doing better in the lab. We are doing about 0.3. Yield is exceeding our expectations. The main workloads that we are getting a lot of traction, it is incredibly, efficient for decode.

Amon argues smartphones are essential for agentic AI and that on-device compute must take a step-function leap to run 100B-parameter models by 2028.

The smartphone is an essential device for agentic experiences as it is the device that knows people best. it is always with them has their contacts, preferences, and permissions. For perspective, leading AI companies are targeting smartphones by 2028 capable of running 100 billion parameter models locally with an all day battery life. Reaching this goal will require a step function change in on device compute performance.

Amon expects optical interconnect industry consolidation, akin to RF front end, to create opportunities, and Qualcomm is talking with laser companies about partnerships.

I think the structure of the industry is going to change. My analogy is what happened on RF front end. And that creates opportunity. So we have been engaged in very interesting conversations with some of the laser companies, especially as different forms of partnerships and strategic partnerships are going to be formed.

Amon says Qualcomm's baseline share at Samsung Galaxy is now north of 70%, up from 50%.

If you think about our business in the past, the baseline was 50%. This baseline is no longer 50%. it is north of 70%, as we have been doing. When you think about future generations of Galaxy. This is how we plan our business and I encourage all of you to plan the same way.

Amon announced that Starlink Mobile picked Qualcomm for a nationwide US terrestrial 5G network, with deployment starting in 2027.

I am very proud to share that StacyX selected Qualcomm to power Starlink Mobile's nationwide terrestrial 5G network in the U.S. Under a multiyear agreement, deployment begins in 2027 and is expected to be 1 of the largest 5G deployments in the United States.

Amon welcomes the jury finding that Arm breached its license agreement and says Qualcomm will keep delivering Oryon CPUs from mobile to data center.

We are gratified that the jury found ARM breached its long standing architecture license agreement ALA with Qualcomm. This outcome is a victory for innovation and competition. And we remain focused on developing and delivering industry leading semiconductor products, including our Oryon CPUs to customers and consumers worldwide from mobile, to data center.

Amon says Qualcomm plans for flat-to-down global handset units in 2027 with the Apple modem transition fully incorporated.

Our plan assumes global handset units will remain flat to down in calendar 2027. The timing and shape of the Apple modem transition is fully incorporated in how we are planning the business and its well understood at this point.

Amon says the fifth-generation automotive platform brings 8x silicon content versus the third generation, with growth driven by share gains.

for cars that you see now on the road that, with Ford third generation platform, the silicon content increase versus the fifth generation is 8x. there is a lot more a lot more compute content. Into the automotive and most of our automotive growth is really share where the design win pipeline translates into revenue.

Amon said fiscal 2027 begins Qualcomm's non-handset revenue inflection, with non-handset revenue projected to grow over 60% and the $5 billion Data Center revenue target reiterated.

“Fiscal 2027 marks the start of our non-handset revenue inflection, including the ramp of our Data Center business,” Qualcomm president and CEO Cristiano Amon said, looking ahead. “Non-handset revenues are projected to grow by more than 60 percent year-over-year, and we are reiterating our fiscal 2027 Data Center revenue target of $5 billion.”

Amon called the SpaceX Starlink Mobile deal a significant win for Qualcomm's infrastructure business and the start of a strategic relationship with great potential.

We are very proud. This new architecture establishes a foundation for next-generation connectivity and network intelligence. This is a significant win for our infrastructure business and the beginning of a strategic relationship with tremendous potential.

Amon said Qualcomm sees sequential improvement in handsets and that Android shipments in China have bottomed.

Chief Executive Cristiano Amon said the company was seeing signs of sequential improvement in handsets and that Android shipments in China had bottomed.

Amon expects Qualcomm to reach $5 billion in data center revenue in fiscal 2027 and says a global hyperscaler is engaged.

Amon said the company expects 5 billion US dollars of data center revenue in fiscal 2027 and has engaged a global hyperscaler.

Amon said Qualcomm's handset weakness stems from constrained supply, notably memory, rather than weak demand, and that more supply could improve results.

We don't have a demand issue; it's a supply issue.

Amon expects Qualcomm's AI data center business to reach USD 5 billion in revenue in fiscal 2027.

CEO Cristiano Amon expects this business to generate USD 5 billion in fiscal 2027.

Amon said Qualcomm's weak results stem from constrained supply, not weak demand, implying more memory supply would have improved the outcome.

"There is no problem with demand, there is a problem with supply," Chief Executive Cristiano Amon said. "If we had more supply, the situation would have been better."

Amon called the SpaceX Starlink Mobile deal a significant win for Qualcomm's infrastructure business and the start of a strategic relationship with great potential.

We are very proud,” said CEO Cristiano Amon of the SpaceX deal. “This new architecture establishes a foundation for next-generation connectivity and network intelligence. This is a significant win for our infrastructure business and the beginning of a strategic relationship with tremendous potential.

Lisa Su said demand for computing keeps growing as applications multiply.

People need more computing to power all these applications.

Lisa Su characterized the market's reaction, including the post-midterm selloff, as short-sighted, signaling confidence in AMD's long-term prospects.

Lisa Su said the market is “short-sighted.”

Su said demand exceeds supply, so supply constraints are the main risk to AMD's 2027 growth guidance.

Su said demand is higher than supply, so if AMD can't build chips fast enough, 2027 growth could fall short of its guide.

Lisa Su said the market can be short-sighted in its jitters about AI demand sustainability, and she sees no signs of caution in AMD's business.

I think the market can sometimes be a little bit short-sighted. It is what it is.

Su argues compute equates to intelligence, so investing ahead of the curve with a three-to-five-year horizon is the right strategy.

What I look at is in the long term, if you believe, which I do, I believe very deeply that computing, in many ways, equates to intelligence. So the more compute you have, the more intelligent your company is, the more problems you're going to be able to solve, the more you're going to accelerate your business. If you believe that foundation, then investing ahead of the curve is the right thing to do because we're always investing for the long term — long term being three to five years and beyond.

Su expects AI and superintelligence to become essential to every aspect of life and business, and believes ecosystem investments reflect this demand.

Over the horizon, I see AI and super intelligence becoming more and more essential to every aspect of our lives and every aspect of our business. So I believe in this demand, and I very much believe that many of the investments that are being made throughout the ecosystem are also seeing some of these same trends.

Lisa Su says AI adoption is moving so fast that formerly static annual plans have become dynamic, which is driving AMD's demand surprise.

with AI and the progress that we're making through AI adoption, uh, throughout the industry, uh, what we're seeing is what used to be, let's call it, you know, fairly static plans, are now actually becoming pretty dynamic. Um, you're not realizing how fast the adoption curve is going.

Su says AMD's data center business will grow 50% sequentially in Q4 and aims to more than double next year, positioning AMD as data center first.

being able to say that my data center business is going to grow 50% sequentially from Q3 to Q4, uh it's a big step up and it's a a testament to how important computing is um in the industry today and you know, we're looking to double the business, more than double the business as we go into next year. Uh we are a data center first company

Su calls the market short-sighted on AI spending concerns and argues compute equates to intelligence, justifying investing ahead of the curve.

I think the market can sometimes be a little bit short-sighted, uh and look, it it is what it is. Um, what I look at is in the long term, if you believe, which I do. I believe, uh, very, you know, deeply that computing in many ways equates to intelligence. So the more compute you have, the more intelligent you are

Su believes AI and superintelligence will become essential to every aspect of life and business, underpinning sustained demand and ecosystem investment.

over the horizon, I see um AI and super intelligence becoming more and more essential to every aspect of our lives and every aspect of our business. So I believe in this demand and I very much believe that uh you know, many of the investments that are being made throughout the ecosystem are also seeing some of these, you know, same trends.

Lisa Su said AMD's server CPU business will roughly double next year, while its data center AI business will grow significantly more than double.

CEO Lisa Su added that its server CPU business will about double, driven by higher unit shipments and higher core counts, while its data center AI business will "significantly more than double."

Lisa Su says that after the midterm backlash against data centers, the tech industry must clearly show every host community the value of AI infrastructure so that everyone benefits.

At the end of the day, what we need to do is ensure that everybody wins from AI. I do think this is an opportunity for us as a tech industry to realize that we need to make sure that the value proposition is clear to every local community that is going to host a data center, and I believe we can do that.

Lisa Su says compute demand is intense, with a very strong order pipeline and demand exceeding supply.

we are in a place Jim where I think there is a um intense need for more compute and when we look at that uh we are seeing a very strong order pipeline. We are actually seeing the uh demand being higher than supply

Su says AMD has landed its first racks at top customers including Anthropic, OpenAI, Microsoft, Oracle and Meta, who are seeing great performance.

we have actually landed our first racks at our top customers including uh anthropic openai Microsoft Oracle Meta and these customers are seeing great performance out of our newest products.

Su expects AMD's business to grow over 80% next year, more than double its long-term model.

We see our business growing over 80% uh next year which is a very significant number. It's over double of our long-term model

Su argues AI adoption is at the very beginning and will require CPUs, GPUs and accelerators together, which is AMD's specialty.

we're at the very beginning of the adoption cycle where people are seeing where you know AI and super intelligence can really take us and for that you need all kinds of compute. You need CPUs, you need GPUs, uh you need all of these accelerators to come together and this is what we specialize in.

Lisa Su says AMD is in very active discussions with its largest customers about the MI500, MI600 and later accelerators.

We’re in very active dialogs with our largest customers on MI500 and MI600 and beyond.

Su expressed confidence that AMD can sustain very strong revenue growth and expanding profitability over the coming years, citing its product portfolio and customer and supply chain ties.

With our leadership product portfolio and deep customer and supply chain relationships, we are confident in our ability to sustain very strong revenue growth and expanding profitability over the coming years.

Lisa Su said demand for Instinct and EPYC keeps strengthening, expects growth to accelerate in Q4 as Helios and Venice scale, and is confident in sustaining very strong growth and expanding profitability.

Demand for our Instinct and EPYC products continued to strengthen, and we expect growth to accelerate further in the fourth quarter as we scale Helios and 'Venice' production. With our leadership product portfolio and deep customer and supply chain relationships, we are confident in our ability to sustain very strong revenue growth and expanding profitability over the coming years.

Lisa Su said PC sell-through slowed late in the quarter as higher memory and component costs weighed on demand.

PC sell-through slowed late in the quarter as higher memory and component costs weighed on demand.

Lisa Su said AMD is planning on the PC market declining by a double-digit percentage in 2027.

Our planning assumption for 2027 is that the market will be double-digit percentage down.

Su says AI agents are becoming the next killer app for cloud CPUs, driving sharp growth in EPYC demand.

Cloud providers also added significant EPYC capacity in the quarter for personal and enterprise agents, which are quickly becoming the next killer app for CPUs in the cloud. Meta's Muse and OpenAI's Dots illustrate how quickly personal agents are growing with more than 10 million people activating Muse agents in the first 6 weeks.

Su raised 2027 expectations: server revenue to double and data center AI to grow faster, with data center above 80% of revenue.

We now expect server revenue to double year-over-year in 2027 and data center AI revenue to grow even faster than we projected 90 days ago. As a result, we expect data center to be more than 80% of our total revenue in 2027 and to drive the majority of our revenue and earnings growth.

Su acknowledges demand exceeds supply and says AMD is working with supply chain partners to expand capacity.

It is true that demand is higher than supply. So we continue to work with our supply chain partners to increase supply. And I do believe that there is a good opportunity for us to do that, just given the strong relationships we have throughout the ecosystem.

Su warns rising component costs will hit consumer PCs, planning for a double-digit market decline in 2027.

Our view is this is one of those times when component prices go up, it affects overall system pricing. One would expect the consumer markets to be more sensitive to that. And so our planning assumption for 2027 is that the market will be double-digit percentage down.

Su says Helios's TCO advantage is the key factor in customers choosing AMD over custom ASICs.

we feel extremely good about where we're positioned in terms of our performance and our total cost of ownership that we're delivering with Helios. That is fundamentally the most important factor when we think about sort of deployment scale and deployments going forward.

Su says HBM and component cost inflation is real and AMD is working with customers to share the costs.

other component costs are also seeing some inflationary pressures in the sense of as there is more CapEx that's put into the overall supply chain environment. We are also working with our customers to include -- to ensure that those costs are shared as we go forward.

Su aims for AMD to capture a larger share of frontier labs' and hyperscalers' compute needs via Helios, MI500 and MI600.

Our aspirations are to be a very significant compute supplier to our largest customers, both frontier labs as well as hyperscalers. And our aspirations are to continue to have a larger and larger piece of their compute needs. And from everything that we see, demand continues to go up.

Lisa Su said AMD expects its data-center server CPU revenue to double in 2027.

it expects revenue from its data-center server central processing units to double in 2027

Altman clarified on X that he wants people to have as much memory as they like in their computers, saying his earlier podcast remarks on the memory shortage were taken out of context.

I actually do want the world to have tons of memory and people to put as much as they want in their computers

Altman says people must always remain in control of the future and OpenAI rejects relegating humans to roles that AI systems decide to reward.

People should always be in control of the future. OpenAI is against the notion that it would be acceptable for people to be relegated to a role of doing things that ‘AI systems think makes sense to reward in humans.’

Altman argues the public should not cede control of the future to superintelligence, and that this should ultimately constrain AI progress.

should not cede control of the future to superintelligence, and that requirement should ultimately constrain AI progress

Altman told staff that if AI visibly goes badly wrong this year, biology is a plausible area for it to happen.

“If something goes visibly really wrong for AI this year, I think bio would be a reasonable bet.”

Sam Altman said he wants people to have plenty of RAM and called for building more capacity, amid severe DRAM shortages.

i want you to have so much ram lets build more

Altman says he wants the world to have abundant memory, that his earlier RAM comments were taken out of context, and that he has long argued for ramping up memory production.

i actually do want the world to have tons of memory and people to put as much as they want in their computers, and it is fucking annoying that this was taken out of context. i have been arguing for the world to ramp up memory production for a long time.

Altman named energy as the single biggest challenge facing OpenAI and the AI industry.

Asked to identify his biggest challenge, Altman answered with a single word: "Energy."

Fouquet said customers are already interested in ordering equipment for 2029-2030 and that ASML's planned 2027-2028 capacity will fall short of demand.

CEO Christophe Fouquet said customers have expressed interest in ordering equipment for 2029 and 2030. In fact, ASML now expects its planned capacity for 2027 and 2028 to fall short of customer demand.

Original (zh)

行政總裁Christophe Fouquet表示，客戶已表達對訂購2029與2030年設備的興趣。事實上，ASML現預期其2027與2028年計劃產能將低於客戶需求。

Haas says Stargate continues strongly with no retreat, and compute demand exceeds supply amid build constraints.

No, not at all. I mean the start the Stargate initiative continues in a very, very strong way. Uh, but data centers are hard to build. Uh, there's a lot of people that are required, uh, there's permitting, infrastructure, uh materials, memory, wafers, chips. Uh and right now, what we're seeing, and I think you've heard this from most of the CEOs you've talked to, uh demand uh for compute exceeds supply.

Haas says Arm AGI CPU orders have exceeded $1B with line of sight to $2B, driven by agentic workload demand.

when we talked about the arm AGI CPU back in March, uh, we talked about a billion dollars of orders and uh, and hoping for more. And then incrementally, we've not only exceeded that $1 billion number, but now we have line of sight to two billion dollars. And that's really driven by demand for agentic workloads, which are CPU bound, CPU driven.

Haas predicts Arm's data center business will surpass its smartphone business next year.

next year, our data center business including our AGI CPU, will be bigger than our business for smartphones. And people didn't think that was possible a few years ago.

Haas is skeptical about space data centers, saying they are unlikely within five years and possible only on a 10-year horizon once power becomes the main constraint.

Uh do I think we'll see data centers in space in two years? No. Three years? No. Five years, probably not. 10 years, maybe so.

Rene Haas says the recent tech selloff reflects short-term post-midterm thinking that ignores AI's large long-term economic potential.

Haas said the recent tech selloff reflects short-term thinking that follows the US midterm elections. In his view, that thinking overlooks the large long-term economic potential of artificial intelligence.

Haas called the post-midterms tech selloff very short-term thinking that ignores AI's longer-term economic thesis.

"The short-term blitz that happened in terms of the election, and people talking about spending, etc., etc.," Haas says on CNBC. "To me, I look at that as very, very short-term thinking."

Haas said intelligence as a commodity is nearly certain to drive GDP growth across the economy.

He says intelligence as a commodity is near-certain to drive GDP growth across the economy.

Haas said Arm is poised to reach 60% data-center share by year-end and its new CPU should benefit from AI agent adoption, with Meta as chief customer.

Haas says the company posted a record second quarter and is poised to exit the year with a 60% market share in the data-center market. The company's new CPU offering should benefit from growing adoption of AI agents like Meta's Muse, he says, adding that Meta is Arm's chief customer for the CPU.

Haas says Arm is highly confident its new Arm AGI CPU will generate $2 billion next year, making data center bigger than smartphones for Arm.

we now have very high confidence in reaching $2 billion in revenue. Just from this product next year, and the result of this will be that the data center business for Arm will become bigger than smartphones.

Haas says Arm's data center share rose from 15% in 2024 to 50% in 2025 and should reach 60% by year-end.

We have gained significant market share in data centers. We had 15% in 2024. 50% in 2025. We think we will reach 60% by the end of this year.

Haas dismisses short-term election and spending worries, arguing intelligence as a commodity will drive huge long-term GDP growth.

In the long term, it is hard to deny that intelligence as a commodity will provide huge GDP growth across all sectors.

Haas argues agentic AI workloads run on CPUs and Arm's energy-efficient processor is well suited, with Meta a major customer.

When all these agents have to plan, track, do all sorts of checks, determine what needs to be done, all this tokenization work is work for the processor. The Arm Aegis processor, because it is extremely energy efficient, is great for this type of task.

Haas said confidence in reaching up to $2 billion of Arm AGI CPU revenue across fiscal 2027-2028 has grown thanks to added supply and customer mix.

Our confidence in achieving up to $2 billion of revenue in Arm AGI CPU across fiscal 2027 and 2028 has increased over the last 90 days, supported by additional supply and customer mix.

Haas expects Arm's share among top hyperscalers to exceed 60% by end of 2026, driven by custom CPU ramps and NVIDIA platforms.

Based upon expected ramps across Google Axion, AWS Graviton, Microsoft Cobalt, and NVIDIA platforms, we now expect that share to exceed 60% by the end of calendar 2026.

Haas argues agentic AI workloads are CPU-bound, which will drive higher core counts and CPU demand benefiting Arm.

What is obvious is that the agentic demands on AI jobs are CPU-bound, and the more agentic workloads that run, that is going to need more CPUs. I think that's what's going to end up driving these higher core count devices, which I think will be very good for us

Haas says Arm's power-efficient compute helps address the projected 38 GW data center power shortfall.

Morgan Stanley projects a 38 GW global data center power shortfall through 2028. Arm is helping customers address this constraint by delivering more useful compute from the power already available.

Haas warns industry supply will be tight and says Arm is working hard to secure supply for very large demand.

I think supply for the industry is going to be tight, and as Jason mentioned, we'll talk a bit more about our future outlook in upcoming quarters. But we are going to be working very, very hard to secure the supply that we need for the expanded demand because we think it's very, very, very large.

Haas notes memory price increases but sees pricing headroom on larger, higher-core-count chips.

memory pricing has obviously increased a lot since we announced the Arm AGI CPU. ... I think it is fair to say that there's probably some pricing elasticity that we could potentially exercise on these larger chips and the chips that have higher core counts.

Haas says the CPU market opportunity is far larger than the $60B consensus, citing $100B by 2030 and possibly $200B+.

when we did the Arm Everywhere day in March, the conventional wisdom was the market was a $60 billion market and was not going to be growing all that much. When we did our calculations, we felt differently, and we talked about a number that day of $100 billion by 2030. I think that number has been increased by various folks. I think Lisa may have even said $200 billion or more.

Haas expects triple-digit data center royalty growth for years as head-node CPUs shift from x86 to Arm.

Google has already announced the conversion of x86 to Axion, connecting to TPU version 8. To Jason's point, this is why we have a high degree of confidence on data center royalties growing triple digits over the next number of years.

Haas indicates an 18-24 month cadence for follow-on CPUs and says Arm designs for rack-level performance, including networking and I/O.

you can think about this market goes at about an 18- to 24-month cadence in terms of production ramps. ... When we think about developing these products, we're thinking about rack system performance.

Haas expects $2 billion in sales of Arm's own CPUs in fiscal 2027 and 2028, with data centers becoming a growing share of revenue.

Chief Executive Rene Haas said the company expects $2 billion in sales of its own CPUs in fiscal 2027 and 2028, stressing that the data center business will account for a growing share of revenue.

Rene Haas says Arm is seeing a huge pickup in demand for all technology related to the datacentre space.

We’re just seeing huge pickup for all technology around the datacentre space,

Haas says Arm now views its $2 billion AGI CPU sales target for fiscal 2027-2028 as very achievable, having secured supply with TSMC and others.

We think the two (billion) is now very achievable,

Haas said Arm is constrained by supply and could do more with additional supply, while expressing strong confidence in reaching its goal.

Arm Chief Executive Rene Haas said: "If we had more supply, we could do more. We feel strongly confident about achieving that goal."

Rene Haas (with CFO Jason Child) says AI is driving unprecedented compute demand, with Arm's IP business and new silicon business, led by Arm AGI CPU, both growing rapidly.

AI is driving unprecedented demand for compute across the cloud, devices and autonomous systems, with Arm being the compute platform of choice. Our business is now driven by two areas, both of which are showing rapid growth: The IP business, where partners increase their investment in Arm technology and deploy record numbers of Arm-based silicon, and our new silicon business, where demand for Arm AGI CPU is extremely strong,

Arm's top executives say Arm keeps gaining data center share, with hyperscaler server share rising from 15% in 2024 to over 50% in 2025 on performance and power-efficiency demand.

Arm continues to gain share across the data center, driven by growing demand for performance and power efficiency as AI infrastructure scales. Data center royalties doubled year over year, and Arm-based server market share at the top hyperscalers more than tripled—from 15% in calendar 2024 to more than 50% in calendar 2025,

Arm's executives expect its hyperscaler server share to exceed 60% by end of 2026 based on ramps across Google, AWS, Microsoft and NVIDIA platforms.

Based on current ramps across Google Axion, AWS Graviton, Microsoft Cobalt, and NVIDIA platforms, we expect that share to exceed 60% by the end of calendar 2026,

Prince expects agent-era micropayment business models (Act 4) to become Cloudflare's largest business within ten years.

I don't know exactly how long it will be before it turns into major sources of revenue, but if you look out in the more distant horizon, if you look 10 years out, I would imagine that this becomes the largest part of Cloudflare's business.

Prince argues agents must compensate content creators because agent traffic requires massive infrastructure.

if agents are going to be 1,000 times human traffic online, that's going to require more bandwidth, that's going to require more servers, that's going to require more things to actually power it, and someone has to pay for it. It can't be a free ride for all the agents that are out there.

Prince says AI outperforms humans at measurement functions, justifying Cloudflare's flatter organization with more builders and sellers.

AI is incredibly good at measuring things. And we have found across our business that in many cases, it's better than what humans can do. AI has bias, but the bias is not correlated to the rest of the organization, unlike what everything else has done.

Prince says Cloudflare has partnered with nearly every major frontier AI lab and security firm hunting for flaws in existing software.

We've partnered with nearly every major frontier AI lab and security company searching for flaws in existing software.

Prince says the Act 4 agent-internet business model line began generating revenue and could become Cloudflare's largest segment within a decade.

Act 4—A product line helping define new internet business models—recognized revenue for the first time this quarter: pay-per-crawl, micropayments, and agent disclosure of intent for lower fees. Collaboration with OpenAI provides page change notifications, helping crawlers reduce costs. A decade from now, this business could become our largest segment.

Prince says AI agents handle measurement tasks cheaply and Cloudflare will hire more builders and sellers, exceeding pre-layoff headcount by year-end.

AI excels at "measurement"-type tasks—for example, Heather from the IR team used an agent to complete the entire team’s previous earnings materials, with token costs under a few hundred dollars. We’re shifting toward hiring more builders and sellers, and by year-end, our total headcount will exceed what it was before the layoffs.

Matthew Prince said Cloudflare has built the development platform for the AI-powered Internet, with annualized revenue passing $3 billion and Q3 revenue up 39% year over year.

We crossed $3 billion in annualized revenue with revenue of $782 million, up 39% year over year,

Matthew Prince says Cloudflare is seeing a fundamental rewrite of the Internet to handle machine-to-machine traffic driven by AI.

we are seeing a fundamental rewrite of the Internet for machine-to-machine traffic.

Zhou Hongyi warns that long-half-life foundational principles and judgment must not be outsourced to AI, or people will be unable to spot AI's mistakes.

长半衰期的基础原理和判断力不能外包给AI，否则无法识别AI的错误。

Zhou Hongyi suggests a company's core competitiveness may no longer depend on headcount but on how many AIs take part in real work internally.

公司核心竞争力可能不再只看员工数量，而看内部有多少AI在参与真实工作。

Zhou Hongyi says enterprises are shifting from training to inference applications, which will drive hundred-fold short-term growth in inference compute that surpasses training compute.

Enterprises no longer frequently train large models from scratch, but instead call AI directly through inference applications to solve real problems; this shift will drive a short-term hundred-fold growth in inference compute demand, exceeding training compute in both scale and growth rate.

Original (zh)

企业不再频繁从头训练大模型，而是更多通过推理应用直接调用AI解决实际问题，这一转变会推动推理算力需求呈现短期百倍级增长，规模和增速都将超越训练算力。

Zhou predicts the AI chip market will split into dual tracks, with international vendors dominating training chips and many vendors competing in cost-advantaged inference ASICs.

Zhou Hongyi judges that the AI chip market will shift from a single-pole to a dual-track structure: high-end training chips remain dominated by international vendors, while multiple vendors will participate in inference chips. Dedicated inference chips (ASICs) will quickly penetrate niche scenarios thanks to cost advantages.

Original (zh)

周鸿祎判断，AI芯片市场会从单极格局转向双轨结构：高端训练芯片仍由国际厂商主导，而推理芯片领域会出现多厂商共同参与的局面。专用推理芯片（ASIC）凭借成本优势，会在各类细分场景快速渗透。

Zhou says the looser chip requirements of inference give China's domestic chip industry a chance at differentiated development.

Inference compute places relatively looser requirements on chips, with lower power consumption and more controllable cost, which gives China's domestic chip industry an opportunity for differentiated development.

Original (zh)

推理算力对芯片的要求相对宽松，功耗更低、成本更可控，这为国内芯片产业提供了差异化发展的机会。

Zhou Hongyi says agents are becoming the main vehicle for AI deployment and that inference compute demand is exploding, surpassing training.

We made two forward-looking predictions: first, agents become the main vehicle for AI deployment, with competition shifting from parameter comparison to real-world deployment capability. Second, inference compute demand explodes, with inference usage and growth surpassing the compute needed for model training.

Original (zh)

一是智能体成为AI落地主角，智能体应用走进千行百业，深入生产和业务系统，行业竞争焦点从参数比拼转向落地能力比拼。二是推理算力需求爆发，人们使用智能体完成复杂行业任务或需求时，需要多次调用大模型进行推理，算力消耗巨大，推理算力使用规模、增速全面超越大模型训练所需算力。

Zhou says an agent system let a 10-person team do in three weeks what a 100-person team needed half a year for.

A 10-person team used the rule-abiding, deterministic agent system we built to develop in three weeks a product that used to take a 100-person team half a year, with coding, testing and delivery fully automated.

Original (zh)

使用我们打造的守规矩、守纪律、能保持确定性的智能体体系，一个10人小队用三周时间开发了过去100人团队需半年才能生产的产品，并实现了编码、测试、交付流程的全自动化。

Zhou says programmers won't be replaced but will shift to specifying requirements and standards for AI.

We used to think AI would eliminate programmers. In practice, programmers remain important; their focus shifts from repetitive coding and testing to giving AI requirements and setting standards for AI output. Those who can transition will grow into 'super individuals'.

Original (zh)

过去我们认为AI会淘汰程序员。实践后，我们发现程序员依然重要，其主要精力从写代码和做测试这类重复性工作中，转而从事两类新工作：一是给AI提要求；二是给AI的成果定标准。能转型的程序员不仅不会被淘汰，还会在与AI协作过程中成长为“超级个体”。

Zhou argues open source is the best way to address AI safety, rather than locking models away.

I believe open source is the best means of solving safety problems. When technology is in the hands of a few, and they refuse to address safety, it creates hidden dangers. Guarding against agent risks is not about 'locking up' models.

Original (zh)

我认为开源是解决安全问题的最好手段。当技术掌握在少数人手中时，如果这部分人拒绝解决安全问题，会带来隐患。防范智能体带来的社会风险，不是把模型“锁”起来，而要强调以人为本、智能向善，安全与普惠并重。

Zhou suggests US AI giants invoke safety to preserve their lead and block China's open-source catch-up.

Some US AI giants with a first-mover advantage have repeatedly cited safety to demand halting AI development. There may be two reasons: they hold core technology and don't want others to have it, and China's open-source AI brings technological inclusiveness that could enable it to overtake the US.

Original (zh)

一些美国AI巨头掌握先发优势，多次以安全为由，要求停止AI开发，这背后可能主要有两个原因：一是他们掌握了核心技术，不希望别人也掌握此类技术；二是中国开源人工智能的发展带来了技术普惠，有可能推动实现对美反超，这是他们不希望看到的。

Hassabis says DeepMind has worried about AI safety since its founding in 2010 and still does, because the technology is general and dual-purpose.

We were worried about safety from the very beginning of DeepMind in 2010, and we still are now, because it’s general. So it can be used for anything.

Hock Tan has said Broadcom's AI semiconductor revenue, about $58 billion this year, should roughly double next year and double again the year after.

The chip designer has said it expects AI semiconductor revenue to generate $58 billion this year, which CEO Hock Tan has said should be expected to roughly double next year and double again the year after that.