AI · GPU infrastructure
How AI compute actually gets built.
Latest
Google signed credit-support agreements to backstop Anthropic's ~$19B, 401 MW lease at TeraWulf's Hawesville campus. If Anthropic misses rent, Google can take the lease—or, in some cases, pay a termination fee for equity in the project SPV. Here's how it works, why the contract talks about TPU pods, and how it fits the new cosigner playbook.
Tom Pechulis, , 7 min read
, 10 min read
AMD's Q3 was the first quarter its Helios racks and MI455X GPUs shipped for revenue. Data center sales nearly doubled to $8.5B, Q4 is guided to ~$17B, and Lisa Su now sees 2027 revenue up more than 80%. Underneath: free cash flow of just $554M, a $3B inventory build, and a $1.1B OpenAI-warrant charge. Here's what the print says, in plain English.
, 12 min read
Nvidia trimmed Vera Rubin's published memory and NVLink bandwidth by 11% to 26% in September, without an announcement, just as the racks hit general availability at Oracle. I checked the archived pages myself. Here's what changed, what Nvidia says, why memory bandwidth is the spec that matters most for serving AI, and what it means for buyers and for AMD.
, 9 min read
Anthropic's amended IPO filing puts a number on its compute commitments for the first time: about $895.4B, with $69.8B due in the next 12 months. Here's what's inside that number, who gets paid (Broadcom tops the list, not Nvidia), why Q3 swung to a loss on pricey short-term compute, and what it means for the neoclouds.
, 10 min read
Nvidia is reportedly in talks to guarantee billions of debt for Nscale's West Virginia campus, where Anthropic signed for up to $44.6B of compute. Here's what a residual value guarantee is, what Nscale's S-1 says about the deal, how much credit support Nvidia has already put on its own balance sheet, and why it matters for the neocloud debt squeeze.
, 8 min read
AEP, FirstEnergy, and Dominion all reported this week. The data-center story has shifted from 'how much load is coming' to 'who pays for the wires.' SoftBank is funding a ~$5B set of Ohio power lines at an 11.35% return for AEP, and my math says that's about two cents per GPU-hour.
, 18 min read
A plain-English deep dive on how GPUs get financed: delayed-draw loans, SPVs, swaps, junk bonds, converts, and vendor money. What actually got more expensive for CoreWeave and the neoclouds in 2026, plus my break-even math on what each point of interest costs per GPU-hour.
, 5 min read
Microsoft and Alphabet both undershot Q3 capex estimates, and neither cut demand. The real story is neocloud delivery slips, Alphabet's TPU inventory build, and a Q4 that has to be the biggest quarter of the cycle.
, 5 min read
An 8-GPU HGX node and a 72-GPU NVL72 rack look like the same idea at different sizes. They aren't. Here's what actually changes in the NVLink domain, the network, power, cooling, and how you plan a deployment.
, 4 min read
How many GPUs fit in a 100 MW data center? A simple, honest model from utility power to GPU count, with the assumptions spelled out so you can swap in your own.
, 4 min read
Vera Rubin NVL72 pulls up to 227 kW per rack, and Kyber is headed further. At 54 V that's more than 4,000 amps of busbar. Why the industry is moving to 800 VDC, what changes in the facility, and the economics behind it.
, 4 min read
Amazon, Alphabet, Microsoft, and Meta guided to roughly $720–745B of 2026 capex. Capex now exceeds operating cash flow at two of them. A look at the numbers, the accounting quirks, and what to watch going into Q3 prints.
, 4 min read
NVIDIA's Spectrum-X Photonics hit full production in August. Why moving the optics onto the switch package matters for power, reliability, and the supply chain, and the CPO-vs-NPO fight coming for Rubin Ultra scale-up.
, 4 min read
H100 rental prices went up in 2026, not down. A bottoms-up model of $/GPU-hr for a GB200 build, CoreWeave's P&L as a case study in depreciation and interest, and why $2.50 vs $3.50 is the line between a dud and a deal.
, 4 min read
Same GPUs, very different clusters. Training is a synchronized throughput problem. Inference is a memory-bandwidth and KV-cache problem, which is why prefill/decode disaggregation is taking over. The math, with Llama 3 70B as the example.
, 4 min read
Meta's Llama 3 run hit an interruption every ~3 hours on 16K GPUs. At 100K GPUs, checkpoint and restart time decide your goodput. A worked model of checkpoint size, bandwidth, and optimal interval, plus what it costs when you get it wrong.
, 4 min read
A worked example of sizing an 8,192-GPU scale-out fabric: rails, radix, oversubscription, switch counts, and the optics bill nobody budgets for.
, 3 min read
GPU supply gets the headlines, but memory is the gating item. SK hynix just posted a 76% operating margin, Micron is booked through 2027, and Amazon raised capex because memory got expensive. What that means for how you plan clusters.
, 4 min read
Quantum-X800 InfiniBand and Spectrum-X Ethernet both ship 800 Gb/s AI fabrics. Ethernet is winning on share, InfiniBand is having a rebound, and the right answer depends on operations, tenancy, and how much you value SHARP.
, 5 min read
The average data center rack still pulls about 9 kW. A GB200 NVL72 pulls 120+. Here's the physics, the facility changes, and the money behind direct-to-chip liquid cooling.