Summary

DeepSeek was the administration’s “starting gun”: Krishnan says America’s AI lead is “very, very small,” even though headlines claiming a few-million-dollar training cost excluded ablations and other costs incurred before the final run. He still credits its KB caching, MLA and reasoning work under inferior hardware, and calls DeepSeek and Qwen the best open-source models today. The strategic read-through is a close race, not a lead America can take for granted.
Krishnan proposes full-stack inference share as a scoreboard for victory: an “America Inc.” spanning NVIDIA/AMD GPUs, OpenAI/Grok/Gemini models and applications should capture as much global inference as possible. He floated the share of all tokens run on American hardware and models; Google had just reported “one quadrillion tokens a month or a quarter,” though he explicitly could not recall which.
Compute buildout is the core industrial challenge: after decades of only 1%-2% US power-demand growth, generation, grids, utilities, permits and data-center construction now form a “tangled spaghetti mess.” The response is “build, baby, build”—easier federal-land permitting, less regulatory friction and efforts to get nuclear construction going—plus electricians, technicians and domestic supply-chain capacity.
Open source is framed as both a strategic weapon and an answer to Chinese soft power, not merely a developer preference. Krishnan argues SB 1047 could have ended US open models, while Guo notes Western companies already use DeepSeek and Qwen; Gil argues closed-model incumbents may have incentives to push the anti-open-source narrative, while Krishnan identifies regulatory capture as one dynamic. Krishnan concedes cyber and bio risks but says Chinese models on “every robot, every camera, every car, every device” would carry catastrophic consequences.
The third pillar is exporting an integrated American stack so allies standardize on US chips, models and applications, reversing what Krishnan describes as the restrictive 200-page Biden Diffusion Rule. He points to the Gulf’s American AI Acceleration Partnership, then puts robotics on an 18-24-month horizon, with policy attention rising over the next 6-12 months as startups already use distilled DeepSeek and Qwen.
AI policy is also cultural policy because models increasingly mediate history, facts and everyday questions. The “No Woke AI in the Federal Government” order requires purchased models to be “truth-seeking” and free of undisclosed artificial ideological bias; Krishnan connects that rule to his Twitter-era view that algorithms can turn curated inputs into national narratives.
Execution is mapped to roughly 90 agency actions and three immediate executive orders on infrastructure, exports and ideological bias, with Krishnan insisting, “There is no plan B.” He rejects the label “technocracy,” says the American worker remains central, and acknowledges an AI “event horizon” beyond which reasonable forecasting may break down—making positioning for multiple possible AI futures, rather than one forecast, the stated objective.

Deep dive

1. DeepSeek erased the illusion of a comfortable American lead

As the senior White House policy advisor on AI and one of the plan’s lead authors, Krishnan traces his policy turn to the UK’s AI debates, where he concluded that senior officials misunderstood open source and startups. After President Trump rescinded the Biden executive order on AI, a few of them were given six months to produce the resulting 28-page action plan.
DeepSeek arrived the weekend before he started at the White House, prompting an urgent leadership briefing: was it genuinely faster and cheaper, and had China trained a frontier model for only a few million dollars? Gil’s corrective was that the headline captured a final run, not the likely hundreds of millions spent reaching it; Krishnan agreed the press had imputed a claim the paper itself did not make.
The remaining signal was still formidable: DeepSeek produced novel efficiency work despite weaker hardware and, alongside Qwen, led open models in Krishnan’s assessment. That made it a “starting gun” for a race whose winner could compound gains across productivity, drug discovery, materials and infrastructure—and then translate that flywheel into drones, autonomous weapons and military scale.

2. Global inference share is both an economic and cultural scoreboard

Krishnan’s proposed measure asks what share of the world’s inferred tokens runs on American hardware and American models. He calls the product bundle “America Inc.”: NVIDIA and AMD at the GPU layer, OpenAI, Grok and Gemini at the model layer, then a broad application ecosystem.
The scale is already difficult to pin down. Google had announced “one quadrillion tokens a month or a quarter, I forget which one”; Krishnan’s point was less the denominator than maximizing US share of perhaps ten quadrillion global monthly tokens.
Gil’s addition — worth keeping: models export culture just as movies and social media did, while becoming trusted sources for history and facts. Chinese omissions around Tiananmen Square demonstrate one failure mode, but he also noted political slant inside US models.
Krishnan compared that power with Twitter’s pipeline from selected accounts into trends, Moments, journalists and “People on the internet are talking about this” stories. The new “No Woke AI in the Federal Government” order therefore requires procured models to be “truth-seeking” and without artificial ideological bias unless its source is disclosed.

3. “Build, baby, build” confronts an energy system not tested for growth

The plan’s three pillars resemble a technology-company strategy: build compute infrastructure, remove barriers to model and application innovation, then make the world adopt American technology and standards. Federal-land permitting for data centers is one immediate target.
Krishnan’s honest answer on required capacity and the first bottleneck was “it’s complicated.” Decades of roughly 1%-2% power-demand growth left generation, utilities and the grid untested, while state incentives and overlapping rules for water, emissions and construction created a “tangled spaghetti mess.”
Asked which energy sources will power the buildout, he declined to quantify the mix and returned to permitting. He singled out nuclear as having been obstructed by the “climate lobby and the doomers,” while stressing that the buildout also needs domestic manufacturing, construction workers, electricians and technicians—not only engineers.

4. Open weights are a strategic weapon with real, but contested, risks

Krishnan argues California’s SB 1047 “would’ve been the end of open source” in the United States. His broader concern is that one state’s rules can become de facto national law for every company operating there, so AI regulation should be handled nationally rather than through a patchwork.
Gil framed open source as an antidote to centralized control: the internet and crypto rely on open protocols, while restricting AI development to three or four companies concentrates power in a small number of companies that could then be controlled by government. Silicon Valley’s advantage is that “anybody, any day” can start.
Guo’s harder strategic point was that open models will happen, and Western companies already use DeepSeek and Qwen. Gil added that model internals remain opaque: code generated today could conceivably contain a conditional payload that activates years later inside critical infrastructure.
On p(doom), Krishnan conceded the need to monitor cyber and bio risks. Gil argued that closed-model companies may have incentives to push the anti-open-source narrative; Krishnan identified regulatory capture as one dynamic. His security rebuttal invoked Linus’s Law—thousands of students and researchers pounding on a 500-gigabyte Hugging Face model may expose problems better than one lab’s small safety team.

5. Export policy and robotics take the stack into the physical world

Krishnan describes the 200-page Biden Diffusion Rule as making GPU exports extraordinarily difficult even to enthusiastic allies. The third pillar reverses that posture: ship American GPUs through arrangements such as the Gulf’s American AI Acceleration Partnership, then use their installed base to pull through American models and standards.
Robotics becomes “super key” over the next 18-24 months, with policy focus intensifying within 6-12 months. Because startups already use distilled DeepSeek and Qwen, the intended response is not only better US robotics companies but competitive American open models that can push American products and standards across robots, drones, vehicles and other devices.

6. The administration is betting that technical fluency can accelerate execution

Gil counted something like 90 agency actions; Krishnan pointed to three already signed orders covering infrastructure, exports and ideological bias. His implementation doctrine was categorical: “Go, go, go. No time to waste. We’re getting it done. There is no plan B.”
The claimed operating advantage is technical fluency inside government. Krishnan cited David Sacks explaining inference, high-bandwidth memory and the shift from pre-training toward post-training in policy meetings, alongside officials’ ability to call industry contacts directly.
Guo provocatively asked whether this amounted to technocracy. Krishnan rejected the label and placed the American worker at the plan’s center, while acknowledging multiple plausible AI timelines and an “event horizon” beyond which he said reasonable discussion of how AI might play out could break down. The objective across those scenarios is to preserve US capacity to capture scientific and productivity gains: “Very simply, we wanna win.”

The RLVR Revolution — with Nathan Lambert (AI2, Interconnects.ai)Why Apple Needs a Management Overhaul & Why Google is Catching Up with Hyperscalers