1/ Most look at data center moratoria and see a power bottleneck. The real story is a structural split in how AI infrastructure gets built.
2/ Voters passed 40 of 41 local data center restrictions this month, many drawing the line at 20-25MW. @antavedissian modeled the fallout: our base case has 41% of 2031 inference running under 20MW across devices, metro pods, solar farms, and old telecom buildings.
3/ Small sites lose on raw hardware cost, so the whole model runs on routing. Utilization has to stay high so GPUs don't sit idle. Scheduling has to respect model state and latency. Verification has to prove the work ran untouched.
4/ Most likely casualty: assumption that AI compute ends up looking like the power grid, owned by a handful of mega-campuses.
5/ Instead of looking like the power grid, inference starts looking like the internet: 10K fragmented boxes acting like a cloud that don't require permits.
2/ Voters passed 40 of 41 local data center restrictions this month, many drawing the line at 20-25MW. @antavedissian modeled the fallout: our base case has 41% of 2031 inference running under 20MW across devices, metro pods, solar farms, and old telecom buildings.
3/ Small sites lose on raw hardware cost, so the whole model runs on routing. Utilization has to stay high so GPUs don't sit idle. Scheduling has to respect model state and latency. Verification has to prove the work ran untouched.
4/ Most likely casualty: assumption that AI compute ends up looking like the power grid, owned by a handful of mega-campuses.
5/ Instead of looking like the power grid, inference starts looking like the internet: 10K fragmented boxes acting like a cloud that don't require permits.
I think total AI inference grows ~1,000x in the next five years. But the power to run it only grows ~10x.
So where does all that AI actually run? I modeled it. 🧵
4
0
8
1,381
I've gone down some weird algorithmic rabbithole and now my feed is mostly Jensen & Alice deepfakes in Bollywood songs. I can't look away and I need to it stop.
9
7
59
4,781
anand iyer retweeted
I think total AI inference grows ~1,000x in the next five years. But the power to run it only grows ~10x.
So where does all that AI actually run? I modeled it. 🧵
So where does all that AI actually run? I modeled it. 🧵
5
3
16
1,940
Formal verification is finally becoming practical. Instead of taking human experts years to mathematically prove low-level code as safe, agentic LLM workflows can now do it in weeks on real-world production systems.
Announcing a major milestone achieved by my team @GoogleDeepMind: we have proven key security properties of the machine code underlying the Signal messaging protocol. This is the first time the actual production code of a widely deployed protocol has been formally verified – by hout AI – and at this depth.
More than a research breakthrough, it’s a prime example of how frontier AI can be a force multiplier to counter the speed and scale of modern cyber threats.
Formal verification offers a higher standard of security; however, it’s extremely complex and time consuming, which is why in many cases it’s considered to be impractical. Similar efforts could take teams decades of combined effort. We did it in a month’s time.
Using AlphaProof Code, Google DeepMind's new agentic framework for formal verification that runs on LLMs including Gemini Carbon, we verified the Signal protocol down to the actual code running on devices. This milestone lays the groundwork for a future in which vital systems, from medical device firmware to power grid software, ship with mathematical proof that their protections work.
Reaching that future will take continued research from our community. In that spirit, we're open-sourcing our formal specifications and verification artifacts so others can build on this work.
On a personal note, this moment is of deep significance to me. I started my own research journey more than two decades ago as an intern working on checking the consistency of complex software systems. Seeing this grand challenge made practical with modern AI is extraordinary.
I'm extremely proud of my team members: Swarat Chaudhuri, Andrew Ferraiuolo, Anton Kovsharov, Tej Chajed, Sam Schlesinger, Karn Seth, Jonathan Katz, and Charles Sutton for taking on this grand challenge and delivering, and grateful to the Signal team for their partnership.
deepmind.google/blog/formal-ve…
3
1
23
2,358
anand iyer retweeted
Agentic payments will soon be a standard feature of mainstream applications.
From shopping apps to banking platforms, companies will want agents that can make purchases, move money, and transact on their customers' behalf.
We built Rain Agent Vaults to give developers the financial infrastructure to make that possible.
Wallets, cards, bank payouts, stablecoin transfers, and spending controls. All through one integration.
Built to fit the needs of each application.
From shopping apps to banking platforms, companies will want agents that can make purchases, move money, and transact on their customers' behalf.
We built Rain Agent Vaults to give developers the financial infrastructure to make that possible.
Wallets, cards, bank payouts, stablecoin transfers, and spending controls. All through one integration.
Built to fit the needs of each application.
Introducing Rain Agent Vaults.
Embed agentic payments into any app.
Your agent. Your brand. Your customers and economics.
Powered by Rain. Owned by you.
13
7
84
3,395
anand iyer retweeted
Maximizing the ROI with reliable and efficient data-center-as-a-service x.com/i/article/2120…
0
2
4
434
anand iyer retweeted
Just 6 months ago roboticists were convinced that text based LLMs could never reason about the physical world and that VLA / world models were the correct way forward.
Fable 5.2 shows substantially improved skills in controlling physical robots, even on tasks it has never explicitly been trained to do.
On RoboDojo, a benchmark with a simulated two-arm robot, it outperformed other models and most AI systems built to control robots.
23
19
536
31,219
anand iyer retweeted
Custom ASICs will dominate (and continue dominating) in the coming years 🫡
We're going to see a huge push from compute-heavy startups and companies to build their own chips, optimized for their workloads (e.g. inference) or products (e.g. satellites, robotics, etc)
We're going to see a huge push from compute-heavy startups and companies to build their own chips, optimized for their workloads (e.g. inference) or products (e.g. satellites, robotics, etc)
I think the key things to watch going forward are
1. The arrival of the ASIC era
2. The arrival of the inference era
3. Who holds the competitive edge in memory, and more broadly, across the supply chain
I’ll put together an article on the bigger picture and share it soon, quite a backlog of articles to write…
Btw @BSPK_ Thanks for painting a clearer picture!!! 😍
2
4
54
2,510
anand iyer retweeted
Clear to see in EDA 🤷♂️
The incumbents' view of AI is orchestrating slow legacy tools. The real unlock is using AI to rebuild the chip design engines themselves. And not just for faster tapeouts, but for searching far bigger design spaces and uncovering novel architectures
The incumbents' view of AI is orchestrating slow legacy tools. The real unlock is using AI to rebuild the chip design engines themselves. And not just for faster tapeouts, but for searching far bigger design spaces and uncovering novel architectures
Marc Andreessen says the gap between what AI-native startups and incumbents think AI can do is "the widest he's ever seen."
His exact words: "The AI-native early stage startups have just a completely different view on what they can do with AI than the typical incumbents in Not slightly different. Completely different.
That's the actual risk for large companies right now. It's not slow adoption — it's a fundamental miscalibration of what AI is capable of. The startups competing against them haven't made that mistake.
The asymmetry compounds fast. Incumbents build roadmaps around a floor. Startups build businesses around a ceiling. By the time the floor catches up to the ceiling, the market has already moved.
9
4
61
5,245
AI makes custom silicon cheaper to start and harder to finish. Writing RTL is now practically free and will be agent driven. So more teams will tape out. More respins.
The more I am looking at AI for semiconductor design, the more it seems to me that:
1. RTL generation is just coding
2. 100% of RTL will be generated by agents
3. This means frontier models will dominate
And to be clear, we still need specialized tools for those agents that do verification, physical simulation, layout etc.
5
4
63
6,487
anand iyer retweeted
Interesting analysis! Disaggregate Inference and heterogenous compute to maximize goodput is the way to go. Addressing the operational complexity and providing operational excellence need to be a part of this story from Day-One!!
Link: x.com/tsait26/status…
Link: x.com/tsait26/status…
A chip that’s 3× faster but idle half the day isn’t 3× cheaper:
A prefill or decode device only pays off when your traffic matches the ratio you bought, and that ratio keeps changing. Your GPUs can switch roles when that happens, but specialized silicon just sits idle while you tware keeps moving the bar too, since megakernels alone have pushed low-concurrency decode on ordinary GPUs up by 40–75% this year, and every one of those gains is another knock on specialized chips.
Consider: “how fast?” but “how fast, times how often it’s actually busy, after moving the KV cache over the network?” If your GPUs run at 70% and their chip averages 40% across a real year of traffic, it needs to be ~1.75× better just to tie. For a single model with very stable traffic (like a premium fast tier), that can work out, but in many case, general-purpose computing can be more useful.
0
1
7
895
The kindest thing a VC can tell a founder is "don't raise from us”. Because not every great company needs to be an all-or-nothing outcome to be a massive success.
6
0
26
3,937
1/ People are reading the Aaronson-Ambainis proof as "quantum is dead." It actually says something much narrower, and more useful.
2/ The claim: if a problem has no structure, a quantum computer can't beat a classical one by more than a polynomial factor. Most theorists already believed this. Now there's a proof.
3/ But every quantum use case that matters runs on structure.
Shor exploits periodicity.
Simulating molecules exploits the fact that nature follows local physical laws.
Neither is a black box.
4/ What it hurts: "exponential quantum speedup for generic optimization"
5/ Quantum's real job was always simulating nature: catalysts, batteries, enzymes, new materials. That's untouched. If anything, the landscape just got clearer.
2/ The claim: if a problem has no structure, a quantum computer can't beat a classical one by more than a polynomial factor. Most theorists already believed this. Now there's a proof.
3/ But every quantum use case that matters runs on structure.
Shor exploits periodicity.
Simulating molecules exploits the fact that nature follows local physical laws.
Neither is a black box.
4/ What it hurts: "exponential quantum speedup for generic optimization"
5/ Quantum's real job was always simulating nature: catalysts, batteries, enzymes, new materials. That's untouched. If anything, the landscape just got clearer.
Today Physical Superintelligence PBC is announcing a proof of the Aaronson-Ambainis conjecture, one of the central open problems in quantum complexity, with major implications for quantum computing speedup.
psi.inc/research/aaron…
Thread 👇
15
20
197
14,907
Didn’t realize @nithyavraman could be my cousin. Shocked my WhatsApp isn’t blowing up more.
2
0
16
3,384
anand iyer retweeted
Industrial arms generate 42% of robot revenue in my dataset, but account for 15% of its attributed value.
Intuitive Surgical alone accounts for 56% of the value.
I looked at where robot makers earn their money and how investors value it in the link below...
Intuitive Surgical alone accounts for 56% of the value.
I looked at where robot makers earn their money and how investors value it in the link below...
3
2
42
3,593
