bruce

Two years ago we started building @Halcyon because we believed that modern AI could fundamentally change how the energy industry accesses and uses information ⚡️

@ShanuMathew93 @DavidCahn6 Yeah agree time for an update for sure … it’s moving fast

Liked that piece a lot when it came out but I think a lot has changed on the ground. Labs started growing faster, cloud businesses are seeing real revenue and returns, and there is far more durability around the agentic coding use case. The Claude code wow moment happened in late-25.

The “$50B per gigawatt” math is a useful shorthand, but it breaks quickly if you treat hyperscaler capex as a clean annual gigawatt tracker. A fully loaded, state-of-the-art 1 gigawatt AI data center campus can plausibly land in the $40B-$60B range once you include accelerators, servers, racks, networking, land, shell, electrical gear, cooling, grid interconnect, and power infrastructure. There’s lever there - cheaper equipment, smaller size, etc. But annual capex does not map neatly to “X gigawatts delivered this year.” The spend is split across very different asset classes. Microsoft said roughly half of recent cloud and AI capex went into short-lived assets like graphics processing units (GPUs) and central processing units (CPUs). Alphabet said about 60% of its Q1 technical infrastructure investment went into servers, with 40% in data centers and networking. Those are not the same economic asset. Microsoft depreciates computer equipment over 2-6 years and buildings/improvements over 5-15 years. Amazon says data centers can have 30+ year useful lives, while chips, servers, and networking gear are more like 5-6 years. The bottlenecks also don’t arrive on the same timeline. Servers can move faster than substations. Transformers, turbines, interconnection, permitting, land, transmission, and utility upgrades can run on multi-year clocks. The International Energy Agency has said procurement can take 2-3 years for cables and up to 4 years for large power transformers. Gas turbine lead times are now reportedly beyond 5 years in some cases. Then there’s inflation inside the stack. Meta raised 2026 capex guidance to $125B-$145B, explicitly citing higher component pricing and additional data center costs. Memory is the current pressure point. That means some capex growth is price, not incremental megawatts. So yes, using $40B-$60B per gigawatt is fine for rough framing. But taking the raw $700B+ hyperscaler capex number and dividing by $50B to declare “14 gigawatts this year” is too clean. The actual answer depends on asset mix, timing, replacement spend, supplier inflation, power availability, and which bottleneck is binding.

@eladgil Same feeling

What a moment of acceleration Breathtaking moment in Silicon Valley (and the world)

@emollick Agree … shockingly fast resolution (think about Anthropic timeline from first model release).

It is remarkable how quickly this market shook out. Anthropic & OpenAI are in business take-off, at least: they have the model development, enterprise deals, compute deals, government & press attention and consumer recognition. Only potential competition is Google (& maybe Meta)

@ShanuMathew93 🎯

There’s a lot of cognitive dissonance in the power debate right now. On one side: data centers are framed as the source of all evil, driving electricity bills higher and consuming every incremental megawatt. On the other: people still haven’t fully internalized what broad-based electrification and decarbonization actually require. Electric vehicles, heat pumps, industrial electrification, reshoring, transmission upgrades, battery manufacturing, air conditioning growth, etc. This was always going to be an enormous power demand story. That’s why I keep coming back to the same point: AI/data centers may be one of the best things to happen to grid investment in decades. For the first time in a long time, you have: -governments prioritizing infrastructure -corporates willing to sign long-duration power contracts at above market prices -capital markets willing to fund generation/transmission -customers that can actually support the returns needed for massive buildouts Does it solve everything perfectly? Of course not. There will be bottlenecks, local stress, permitting fights, bad projects, longer thermal runoff period,and periods of volatility. But absent another generational supercycle where incentives align across utilities, policymakers, infrastructure developers, and private capital, this may be one of the strongest opportunities we get to materially modernize and expand the grid, ever! Also a little convenient how some of the loudest proponents of gas stove bans, EV mandates, and broad electrification are suddenly treating data centers as the sole culprit behind higher power prices.

@ShanuMathew93 @DavidCahn6 Yeah agree time for an update for sure … it’s moving fast

Let’s go ⚡️🤓

Speed to Power ⚡️🤓 h/t @schrep halcyon.io/blog/machine-r…

@eladgil Reasonalbe

TIL the Turing award looks like this 🤯

@duncancampbell Doesn’t scale

It’s depressing how the default business now is essentially a scam. Just pump the crazy hype with no real plan for an actual outcome.

@_sholtodouglas Long Anthropic but short terawatts in orbit …

x.ai/news/anthropic… Incredibly excited to partner with SpaceXAI on Colossus 1 - and looking forward to scaling to many terawatts in orbit :)

@DKThomp Meh model means low usage means low utilization means burning money means open to a deal …

Wow. The AI supply crunch is real. Frontier labs are desperate for compute. Musk has compute capacity but a meh model, and Anthropic has a fantastic model with weak capacity. Now I wonder if Elon continues to refer to his new business partner as “woke AI,” “Misanthropic,” etc.

@benthompson 🎯

It was always clear that Anthropic would get compute if they needed it. Demand finds supply.

@JigarShahDC That’s the number I’ve been hearing is the all in cost Cc:@DavidCahn6

This translates to $50B per GW of compute?

@distinctalerts @Jessicalessin And yet it’s hard to see how they do at the same time 🤔

@boo @Jessicalessin Would be nothing short of a disaster if they couldn’t meet the spend commitments I’d suppose

Good lord. Half-ish of the cloud backlog at Microsoft, Oracle, Google and Amazon is OpenAI and Anthropic????

@distinctalerts @Jessicalessin Depends on how those growth rates sustain or not … nobody can know for sure (spend commit not gtee)

@boo @Jessicalessin Are they generating enough revenue to pay for it all though?

@RichardMeyerDC Yeah … it’s an important reminder.

Just a reminder to Energy Twitter that the reports of xAI GPU feelt is only at 11% "utilization" refers to the model FLOPs utilization during training throughput. That's not the same as overall GPU utilization. And it's not the same as power/energy requirements!

@duncancampbell Banger 💥

The population of NYC requres roughly 35,000 square miles of farmland to support it. That's over 100x NYC's land area.

@pandemona Let’s go 💥

Podcasts are the largest source of untapped data on the internet. But searching audio is still broken. Today, we're launching a product to fix that. Meet the Particle Podcast Intelligence API. Track mentions or topics, get shareable clips, uncover ad intelligence, and get access to rich metadata. Tell your agent: "Go to docs.particle.pro" 🛠️