AMP's Anjney Midha: only 15% of US AI compute is actually being utilized — that's a national security crisis

US AI compute is 85% wasted. Anjney Midha calls it a national security crisis.

Anjney Midha, a partner at Andreessen Horowitz and a builder of the AMP compute infrastructure platform, argues that America's real AI bottleneck isn't chip supply — it's utilization. Only 15% of every dollar of compute capacity in US tenant data centers is actually being used to run models. That number is the core of his pitch, and it shapes how he thinks about the competitive threat from China.

The waste compounds across two layers. For every dollar of long-term lease an AI lab signs today, roughly 40% of available flops disappear in bad scheduling — nodes simply not allocated to jobs. Of the flops that do reach a chip, model flop utilization (MFU) sits below 20%, because chips sit idle waiting on memory, storage, networking, or adjacent chips to hand off work. Compound those two losses and you arrive at the 15% figure. Midha describes the remaining 85% as a national security crisis — the gap that allows China to stay competitive without matching US chip capacity.

His argument against pure scaling is direct: China can coordinate industrial-scale compute build-out in ways the US permitting and financing system cannot match. Beijing can add 100 gigawatts of new energy capacity in a decade without blinking. The US response, in his framing, has to be efficiency — what he calls "output maxing," measuring capabilities and business results as output divided by unit of input rather than raw spending.

“For every dollar of long term lease that an AI lab takes today, you lose about 40% flops in just bad scheduling. The nodes are just straight up not allocated. Then within the chip, you only have about 15% of the chip being utilized. That's called MFU model flop utilization. As a result, if you compound those two things, of every dollar, only 15% of flops are actually being utilized. That 85% — that's a national security crisis.”

Supply constraints and nuclear

AMP's infrastructure arm, which Midha says is being spun out as an independent entity (name not yet announced), is targeting ~2 gigawatts of US compute capacity. The supply chain problem is concrete: AMP is already procuring energy and sites for 2030 capacity because the pipeline is backed up that far. New sites are scarce enough that nuclear is not optional.

Midha introduces Seth Cohen, cofounder of Mavria — an AI infrastructure company AMP has invested in — as a former DOE nuclear policy chief and DOGE operative who worked on permitting for new US nuclear projects. Cohen's contribution on the nuclear side was a narrative shift around spent fuel: moving from Yucca Mountain as the only politically acceptable waste site to a model where 27 governors applied for a new program called the Nuclear Lifecycle Campus. A small number of states are now actively competing to host the full back-end fuel cycle. Cohen frames the 100,000 tons of commercial spent fuel sitting on reactor pads as an asset — containing roughly four times more energy than Saudi Arabia's proven oil reserves — rather than a liability. Mavria remains in stealth on its core product.

Reinforcement learning as the next revenue leg

On where AI token consumption goes after coding, Midha points to reinforcement learning at industrial scale. RL, in his framing, is moving from bespoke craft to repeatable, productized post-training: take a base model, fine-tune on a small, high-quality dataset with clear reward signals, and build a specialized capability very quickly. Wherever formal verification is possible — unit tests in code, X-ray diffraction in materials science — the feedback loop tightens and capabilities compound fast.

His portfolio example is Periodic Labs, which operates a 40,000-square-foot facility in Menlo Park pursuing room-temperature superconductors. The loop runs entirely on RL: AI models predict new material candidates, robots synthesize them, and X-ray diffraction machines test whether the superconducting properties match the prediction. The signal is unambiguous — physics either confirms it or doesn't — and that clarity drives rapid capability gains. Midha says progress there over the past six months has been extraordinary.

AMP structure

AMP runs two distinct businesses. AMP Foundry is the venture capital arm. The infrastructure arm — being spun out separately — is building and procuring the physical compute capacity. Midha confirms AMP has a significant position in Anthropic, which he describes as having grown from zero to well north of $40 billion in run rate, with most revenue coming from everyday customers using Claude for coding rather than from circular partner arrangements. He uses that trajectory to argue that private credit markets need to develop better frameworks for underwriting AI startups as infrastructure counterparties — the financing gap is one of the structural bottlenecks holding back US compute build-out.