From kilowatts to tokens: introducing DesertGrid
If cheap power is the moat, what do you build on it first? DesertGrid is our answer: an OpenAI-compatible AI API on Liwa's $0.10/kWh power, with flash-class models from about $0.15 per million tokens.
Stargate UAE: inside the 5 GW AI cluster rising in the Abu Dhabi desert
Five gigawatts on ten square miles of sand, NVIDIA, OpenAI, Oracle and G42, with the first 200 MW due in 2026. When the giants pick the desert for the biggest cluster outside America, what are they telling you about where the power is?
TSMC, CoWoS & HBM: the real bottleneck behind the GPU shortage
It was never just the wafer. Advanced packaging is booked 52-78 weeks out and NVIDIA reserved more than half of 2026-27. If silicon is rationed upstream, what's the scarce thing downstream?
SpaceX's Terafab: a $119B bet to vertically integrate AI silicon
Tesla, xAI, SpaceX and Intel, one terawatt of compute a year, under one roof. If the people who build rockets now want to build the chips too, what does that say about where the value really sits?
Co-packaged optics: the fabric for million-GPU clusters
A million-GPU cluster's hidden enemy isn't the chips, it's the wiring. NVIDIA's co-packaged optics use 4x fewer lasers for 3.5x more power efficiency. Every layer of AI is being redesigned around power, so where do you want to be standing?
The water question: how do you cool AI in a thirsty world?
A large data center can drink as much water as a town of 50,000. As usage soars and backlash spreads, the desert looks like the worst place to build, until you see it forces the water-frugal answer everyone will need: a closed loop.
Data centers in space: physics, hype, and the $39B question
An H100 already ran an LLM in orbit. Google radiation-tested a TPU for a 5-year mission. Free sunlight, free cooling-by-vacuum, until you do the heat math. Does the orbit actually pencil out?
800 volts DC: rewiring the rack for the megawatt era
As racks climb toward a megawatt, the copper busbar hits a wall of physics. NVIDIA's answer is to rewire the rack at 800 volts. It sounds like a footnote. It's a redesign of the whole building, from grid edge to chip.
Is it a bubble? The trillion-dollar AI build-out, examined
The biggest companies on earth will spend close to a trillion dollars on AI this year, and a sovereign fund just warned of a 35% hit if it sours. Forget bull or bear: when the dust settles, who is still standing?
Blackwell, decoded: GB200, GB300 and the 120 kW rack
72 GPUs, 13.5 TB of HBM, 120 kilowatts in one cabinet you cannot cool with air. Blackwell didn't just raise performance, it changed the physics of the room. Is your facility ready for it?
NVIDIA's Vera CPU & Vera Rubin: the real cost, the real timeline
Full production was declared at CES 2026, first racks ship Q3. But a single rack's bill of materials is now estimated near $7.8M, and memory did something wild. What are you actually buying?
xAI's Colossus: the gigawatt sprint that outran the grid
The Memphis grid offered xAI 8 megawatts; the cluster needed hundreds. So it rolled in dozens of gas turbines and powered itself. Colossus isn't a story about GPUs. It's about what happens when ambition outruns the grid.
Google's TPU vs NVIDIA's CUDA: is the moat cracking?
Ironwood claims ~44% lower TCO than a GB200 server, and Google is making millions of TPUs. NVIDIA still holds ~81% of the market. So why hasn't the dam broken, and what would it take?
The DeepSeek shock, one year on: efficiency made compute cheaper, so we used far more
A year ago it erased nearly $600B of NVIDIA in a day, on the logic that efficient AI needs fewer chips. Demand only grew. A 160-year-old paradox explains why, and why it points straight at the price of a kilowatt-hour.
Nuclear for AI: the gigawatt power deals, and the simpler alternative
Microsoft restarted Three Mile Island; Meta signed 6.6 GW of nuclear. The scramble to buy reactors confesses what the real bottleneck is. But a reactor is years away, and cheap firm power already exists somewhere.
Inside NVIDIA's partner program: how GPUs actually get allocated
NPN, NCP, "Reference Platform", the alphabet soup that decides who gets silicon and who waits. If allocation is gated on the facility, where does that leave a newcomer with racks but no badge?
The US-Gulf chip corridor: why the UAE can suddenly land the best silicon
A Biden-era export rule was torn up and a corridor opened: hundreds of thousands of NVIDIA's best chips can now flow to the Gulf. If silicon can finally land in the UAE at scale, where should you build?
HUMAIN and the Gulf's sovereign-AI race
Saudi Arabia built a state AI champion and signed NVIDIA, AMD and Cisco, aiming at 6 GW by 2034. The Gulf isn't buying AI, it's building the factory. So who serves everyone the champions can't?
Anthropic's gigawatt: when compute is measured in power plants
Anthropic didn't announce a chip count. It announced a gigawatt, up to a million TPUs online in 2026. When a model lab measures itself like a utility, the unit of AI ambition has quietly become the watt.
OpenAI's AMD bet: 6 gigawatts and a path to 10% of AMD
OpenAI agreed to 6 GW of AMD silicon and took a warrant for up to a tenth of AMD. When the biggest buyer becomes an owner to crack the monopoly, why would your facility still bet on a single chip?
Stargate's $500 billion question: who actually pays for the power?
$500 billion, 10 gigawatts, the largest infrastructure pledge ever made. The chips get the headlines. But the power bill is the number that compounds for a decade. So who actually pays for it?
Rubin CPX: NVIDIA splits inference into two kinds of chip
NVIDIA split inference in two and built a different chip for each half. The named use case for the new one? Generative video. When silicon specialises by workload, can a generic facility still host the result?
Meta's Hyperion and Prometheus: a data center the size of Manhattan
Zuckerberg measured his next data center against Manhattan. Not a building, a borough. When compute is sized in boroughs and powered by self-built gas plants, the scarce inputs are land and power, not chips.
More in the series, AMD Instinct, the memory wall, neoclouds, the power wall, liquid-cooling supply chains, sovereign AI in MENA, and the TCO of colocation vs cloud. Want one written next? Tell us.