The Weekend DeepSeek Dropped, the White House Called

Sarah Guo and Elad Gil welcome Sriram Krishnan as senior White House policy adviser on AI and a lead author of the American AI Action Plan released that week. The joke title is “your excellency”; the real résumé is Silicon Valley. He did a tour of the large consumer social companies, then spent four years at Andreessen Horowitz, including a stretch competing with both hosts for Series A term sheets. For all of that time, he says, he had no real intention of joining government and was not particularly interested in policy. He did not even know what the word meant.

The turn came when he moved to England a couple of years earlier to run a16z’s international effort, just as the UK became a hotbed of AI policy. The Bletchley Park safety summit sat at the peak of the effective-altruist versus e/acc fight, and he was pulled into the rooms. What struck him was that people in very senior roles, in the United States and elsewhere, did not know what they were talking about. On open source and on helping startups, he was convinced they were doing the wrong thing, in a way the industry itself did not appreciate until much later. That is when policy became interesting.

After President Trump was inaugurated, the first week produced two acts Krishnan still treats as the brief. One rescinded the Biden executive order on AI, which he calls bad and awful. The other said America should dominate and win. A few people, including him, were told to come back in six months with a plan for how. Everything since, he says, culminated the day before this taping: a 28-page action plan and a bundle of executive orders. The document is meant to read like a technology company’s strategy memo. He wants people in the industry to go read it.

The catalytic moment was the day before he started the job. DeepSeek landed that weekend — there was even chatter, he notes, that China had timed it for the new president’s swearing-in — and the White House asked him to come brief leadership. What is this. Is it cheaper, faster, some magic way of training a model for a few million dollars instead of hundreds of millions. That was the narrative of the weekend, and he and David Sacks walked the building through it. He now calls it a starting gun.

At that moment DeepSeek was the only reasoning model that was not OpenAI. Krishnan does not think Claude or Google had one out yet. It sat high on the leaderboards. The cost claims were murky. Gil reminds him how overstated they turned out to be: a few million for the last training run, with hundreds of millions likely spent on the path to get there. Krishnan agrees the headline was inflated, and he is careful to add that the paper itself does not claim a few million as total cost. That figure, he says, is what the press imputed.

The work still gets credit. DeepSeek did not have hardware as good as the American labs, and it still produced real technical pieces — KV caching, MLA, and, depending on the theory you like, chain of thought found in-house or with some help from American companies. Symbolically it said the lead cannot be taken for granted. They were the best open-source model then, and even now he would probably say the Chinese models, DeepSeek and Qwen, still are. The country is in a race, and a close one.

Guo asks what it means to win, and what losing would look like. Krishnan starts from a claim he suspects the hosts share: AI may be the most transformational economic and cultural force of their lifetime. Whoever gets ahead gets a flywheel — productivity, drug discovery, new materials and technologies feeding back into infrastructure and the economy — the network-effect logic Silicon Valley already understands. The military version of the same picture, from drones to autonomous weapons, is why he does not want another country to reach that scale first.

The measurement problem is the interesting part. Google had just announced on the order of a quadrillion tokens of inference; he forgets, on air, whether the unit is a month. His working model is: suppose the world infers some number like ten quadrillion tokens a month. What share of those tokens runs on American hardware and American models, and how do you maximize it. He describes the country as America Inc., a product stack from GPUs at Nvidia and AMD, through a model layer at OpenAI, Grok, Gemini and others, into an application layer of agents and software. The policy job is to make that stack dominate token share.

Gil adds the piece the stack talk can miss: cultural export. Movies did it, then social media; now people treat models as a source of truth for history and facts. Chinese models that omit Tiananmen are the famous case; some US models, he notes, have their own political slant. Krishnan takes the point personally. He grew up in India; a lot of his exposure to Western culture arrived through an American internet. Imagine, he says, if 1995’s internet had been run by an adversary. DeepSeek, when it landed, was full of material that does not align with American values. That is not a side issue. It is one of the things the new orders try to address.

The spicy order, as he warns the audience, is titled along the lines of no woke AI in the federal government. From day one the administration has been fighting DEI, “wokeness,” critical race theory, whatever the label, inside the government. The rule for procurement is simple. Models bought with taxpayer dollars have to be truth-seeking, and they cannot have artificial ideological bias added; if bias is added, say where it came from. That is meant to cut against a model that says nothing happened in Tiananmen in the early 1990s, and against the other culture-war examples in the order: a Pope generated as a Black person, misgendering ranked worse than a thermonuclear explosion. These systems will sit inside government workflows. At least make the bias visible.

Guo, knowing his years inside social products, asks where the old platform-versus-publisher fight applies to models. He spent a long time at Facebook, now Meta, and at Twitter, and what stayed with him is how easily cultural bias could be injected into algorithms. Pick the right accounts, feed the trending system, land in Twitter Moments, and the next morning every journalist and editor is awake. BuzzFeed writes that people on the internet are talking about this. He watched it happen over and over. It left him with a blunt line: Twitter, or X, is the memetic battleground of the ideological wars.

He expects AI to rhyme. His kids already ask ChatGPT everything, from history and geography to silly children’s questions. It is easy to imagine someone injecting their own cultural biases into that layer. The executive order’s examples are there to say the government does not want systems at the center of its work artificially loaded with an ideology, at least not without transparency. That is the cultural half of the stack argument, before the plan turns to concrete pillars.

David Sacks and Krishnan tried to write the plan so it would make sense as a tech-company strategy. If America is going to win a race with China, it has to do three things. The first is infrastructure. Scaling laws demand compute and data, and in the United States the grid and the permitting around new data centers have made projects hard to get off the ground. The president’s slogan is “build, baby, build,” a play on “drill, baby, drill.” One concrete move is easier permitting for data centers on federal land, cutting through older environmental laws and other rules that get in the way. Power the models as they scale. That is pillar one.

Pillar two is innovation: let companies that exist, and companies that do not exist yet, build applications and models as fast as they can. The first cut is red tape. A year and a half earlier Krishnan was still in California when the state nearly passed SB 1047, which he says would have been the end of American open source — no Llama, and no models like it. Other states want their own versions. He wants AI treated as a national priority, regulated at the federal level, because a small state’s rule can become the de facto law of the country. Companies with an office there simply apply the strictest version everywhere, the way EU rules radiate. Series A and Series B firms should not have to play that game.

Open source is the other half of the innovation pillar, and one of the reasons he entered policy in the first place. The Biden administration, in his telling, tried to scare people about how unsafe open weights were; SB 1047 tried to ban them in many ways. The new order says this is a space the United States needs to win, and it points to resources for research. Kids in dorm rooms, startups, and anyone who wants cheaper inference on an IoT device or a robot already live there. Gil notes that much of the internet already runs on open software and open protocols. Pulling open source out of AI would centralize power in a few firms the government can control. Krishnan says that is the talking point, and it is right. Biden-era AI policy looked like control of three or four companies that submit models for testing. DC thinks in centralization. Silicon Valley does not, and he would rather enable a Y Combinator seed than concentrate power inside a ten-mile radius of where he is sitting.

Pillar three is a reversal of the Biden diffusion rule, a 200-page document that, as he summarizes it, made it extremely hard — “illegal,” in his shorthand — for America to export GPUs, even to enthusiastic allies. Jensen Huang and Lisa Su could not get chips out. The new approach is to get the world onto the American stack. Gulf deals under the American AI Acceleration Partnerships are the early case: send GPUs, and you probably get them running American models rather than someone else’s. Robotics has its own section in the plan because Krishnan thinks the domain gets “super key” in the next 18 to 24 months. When he talks to robotics startups he already sees a lot of distilled DeepSeek and distilled Qwen. The response he wants is an American open-source alternative, pushed as the standard, in the US and abroad.

Asked, as a systems person, what the first bottleneck is — grid, generation, workforce — Krishnan says the whole system was never battle-tested. Someone showed him a number: US power usage grew about 1 to 2 percent for a very long time. Gas turbines, coal, renewables, a regulatory freeze on nuclear, state utilities with little incentive to innovate, a grid that was never pushed, and a patchwork of water, emissions, and other environmental rules. It was tangled spaghetti, and until two years ago that was fine, because nobody was using dramatically more energy than a decade earlier. Then scaling laws arrived. The attack is at every step: generation, easier data-center construction, the right energy sources lined up. An infrastructure executive order went out the day before the interview. The National Energy Dominance Council works with Interior Secretary Burgum and Energy Secretary Wright. The short answer, he tells Guo, is that it is complicated, and there will be more.

Gil’s industry theory is that fiber is cheap to lay and the electrical grid is not, so data centers will cluster on cheap power and run fiber in, rather than chase telecom capacity. Krishnan’s version of the government’s job is narrower: get the red tape out, make permitting easy, and restart nuclear, which he says the climate lobby and the doomers blocked for decades. Manufacturing sits next to energy. America needs not just engineers but electricians and technicians, construction, and the ecosystem back in the country. The action plan, he says, has a bunch of language on that stack of jobs.

Gil then asks the political question: about 90 agency actions are listed. How do they become industry behavior, and what is Plan B if they do not. There is no Plan B. This administration, Krishnan says, moves fast — executive orders in the first week, three more signed the day before this conversation, one each for infrastructure, export, and ideology. They are already at work on pretty much all of it. Industry reaction has been a flood of people excited that the government might actually understand AI and wants American companies to build. Go, go, go. No time to waste.

Gil’s observation is that prior administrations, of both parties, often hired nice people from tech who were not the core of the industry. This time the caliber is higher, and it shows. Krishnan points to Emil Michael as under secretary for research and engineering, with DARPA in the Pentagon, and to David Sacks in meetings walking through how inference works, what high-bandwidth memory is, and how the field has shifted from pre-training to post-training. Washington, in his earlier diagnosis, suffered from a lack of real technical understanding. That, he thinks, is what this cohort is there to fix.

Guo asks, a little as a trap, whether the United States should simply be a technocracy. Krishnan will not take the word. America has the leading technology ecosystem in the world, it is in an intense and still-close race, and the job is to protect, preserve, and extend the lead. At the heart of the action plan, though, is the American worker. On stage the president had listed inventions the country already gave the world: the integrated circuit, Shockley and the transistor, Fairchild, the internet, PageRank and Google, the iPhone in Cupertino. The question is how to keep that lead in AI. He has read Daniel’s *AI 2027* and much more optimistic takes. There will be an event horizon beyond which reasonable discussion of the path gets hard. In any of those scenarios he wants the United States positioned to take the productivity and science, and to be set up for whatever comes next. Put more simply, when Guo presses the national-interest version: we want to win.

The last hard question is the strongest case against open weights: that p(doom), or some cycle of abuse, rises when powerful models are released. Krishnan says the action plan already talks about systems to identify cyber risk and bio risk. The difference from the Biden years, and from people who talk about p(doom) a lot, is temperament. He points to the vice president’s Paris speech: embrace AI with optimism rather than fear. Two fears got mixed together. One is models that do terrible things. The other is that open-sourcing is giving secrets to China. DeepSeek, he says, refuted the second on contact. China was building the models just fine by itself; it was American models that were behind.

On p(doom) itself he grants a fair question that needs vigilance. Then he puts the race back on the table. There will be catastrophic consequences if Chinese models run on every robot, every camera, every car, every device, and that reality has to be faced. Gil adds the incentive problem: a lot of the loudest p(doom) push comes from one or two large closed-source companies, classic regulatory capture, the old pharma move of working with government to keep new entrants out, sometimes wrapped in concern for humanity. Krishnan agrees capture is in the mix, along with effective-altruist schools of thought. His rebuttal is older than either. Open source on the internet has defaulted to safer. Linus’s law says more eyes make every bug shallow. Twenty years of security practice says the more scrutiny you put on libraries and browser engines, the safer they get. A 500-gigabyte file on Hugging Face, pounded on by thousands of students and researchers, has a good chance of surfacing issues better than a small safety team inside a large lab. That is why he remains a fan of open source, sometimes as the more secure option, not the less.