The $5.5 Million Was Only the Last Mile

Sarah Guo and Elad Gil open the 100th episode of No Priors with the news of the hour, which by now is the news of the month: DeepSeek. Gil's first cut is that the story is important in some ways and, in others, exactly what a trend line would have predicted. Interest clustered around three things. It was a state-of-the-art Chinese model, open source, that seemed to have caught up on reasoning. The paper claimed a cheap final run, on the order of five and a half million dollars. And a wider narrative asked who was really behind it — a hedge fund, or something else — a layer of mystery that may or may not be real.

On the model itself, Gil is complimentary. Read the paper, he says, and you find novel reinforcement-learning techniques that other labs are already starting to adapt. Similar ideas had been circling elsewhere, but the engineering is real. On cost, everyone he has talked to who is close to the work treats a final run at this scale as a $5 to $10 million event. The interesting question is what was spent before they distilled down to the smaller model. His sense is that people in the know put the compute bill leading up to it in the hundreds of millions. From that angle, the roughly 20 percent drop in Nvidia as the news spread looks, to him, a bit unwarranted.

Guo is struck by the delay. DeepSeek V3, the base model pretrained on internet data to predict the next token, came out in December. Nvidia did not crash on that news. People do not want raw next-word likelihood streamed at them. What they want is post-training: human feedback, high-quality prompt-and-response examples, the instruction fine-tuning that turned GPT-3.5 into ChatGPT. R1, released as a parallel to OpenAI's o1, was the equivalent breakthrough in public understanding. Underneath the product moment sat twenty years of a story in which America leads and China chases. This one showed that Chinese models can get there fast.

Guo agrees that cost was a huge part of the reaction, and that it may have been misstated or misunderstood. She is not even sure final model runs at scale really sit in that price range. What she is sure of is the multiple. Tooling, data work, experimentation, the pretraining run, data generation, post-training, and inference — she is sure she is missing a line item — make it very unlikely that the total spend is only a small multiple of $6 million. Experimentation, as Gil said, tends to be a multiple of the last run. The sticker in the paper is the last kilometer of a marathon.

There is still a real narrative violation, and she names it. Even at a multiple of six million, you do not need a multi-billion-dollar entry price, and you do not need a Stargate-sized check, to sit at the table. Training a GPT-4-level model today is a massive drop from two years ago. That falling curve, more than one Chinese lab's press cycle, is what shook the market. Gil's team ran the inference version of the same arithmetic: in the last eighteen months, cost per token for a GPT-4-level model fell 180 times — 180x, not 180 percent. Training equivalent models has been diving too. DeepSeek, on his read, is roughly on trend, a bit better, with advanced techniques that are real, and still overstated as a rupture.

Gil asks whether cheaper capability is proof that models are commoditizing. Guo points to Artificial Analysis, a site that does not take papers at face value: the operators re-run benchmarks themselves across scientific reasoning, math, coding, multilinguality, and cost per token relative to performance. Eighteen months ago the gaps were enormous. Now the graph is converging, even if some models still lead in pockets. A breakthrough could open the spread again. On average, the market is closer than it was a year ago.

So what is the value of standing at the frontier? Gil lists three or four kinds. Market share and habit: people stay because their prompts and tooling are fitted to you. Using the strongest model to build the next one — data labeling, synthetic data, post-training, coding tools. And a speculation he will not cash: as you near some form of liftoff, the more state-of-the-art the model, the faster it bootstraps the next one, and the lead snowballs. He does not know if that is true. Guo extends a piece she thinks the market has underpriced. A base model high-quality enough for synthetic data generation is a big leveler. If more powerful bases keep showing up, including open ones, the bar for self-improving models keeps falling. You do not have to build the wheel from scratch.

One more consumer fact sits next to the charts. The DeepSeek mobile app became a top contender in the App Store. Guo sees two stories. One is that the cheapest, most capable model on the market actually matters to consumers and they can tell. The other is that the drama was interesting, and people wanted to see whether the leading Chinese model was as good as OpenAI and Anthropic. She believes leading capability can create novel product experiences that draw attention. In this case, she thinks the latter story did more of the work.

The same week, OpenAI shipped Deep Research and, with Donald Trump at the White House, announced Stargate, a massive set of AI-infrastructure investments. Gil wants everyone to try Deep Research. The biggest deal, for him, is that it immediately raises the bar for knowledge work. Work you might once have hired a median intern or analyst to do now has to clear what this product can do, and the comparison is hard. He expects others to copy the pattern, and he gives the team credit. On first use it is more valuable to him in domains he understands less: surveying a field, getting a comprehensive view, finding the experts.

In domains where he has depth, he takes issue with its implicit ranking of authority and its feel for what on the web is good. You will have to audit the outputs. It will orient you, but you cannot take many of the claims as given. That is the AI form of Gell-Mann amnesia, a phrase Michael Crichton coined after the physicist Murray Gell-Mann, who won a Nobel for quarks. You read a newspaper page on a subject you know and find it dumb. You turn the page to a subject you do not know and assume they got it right. Why would they, if they just got your field wrong? People never apply that lesson. They assume the next page is fine.

The psychology has consequences, because these systems are becoming a lot of people's primary source of information. They overlap search in deep ways, and the sources have traditionally been less visible than a blue link. Agents make it worse: they go and do things, then report back, and you do not even know what they read. Gil finds that dangerous from a propaganda and censorship angle. Social networks were version one or two of information pollution. This is the large version — Google mixed with Twitter mixed with Facebook mixed with every media outlet, in one box you interrogate. Whoever controls the output holds a power that is both useful and unsafe. That is why he is glad we live in a multi-company, multi-model world, and why open source, if you care about civil liberties, becomes a load-bearing piece of the stack.

Guo hears several questions packed into Stargate. How much does access to the largest cluster still matter in the race? Is the money real? How deep are the capital markets? What does sovereign and quasi-sovereign capital in the mix mean? She skips whether the checks will clear, and she has no strong view on the last two. The question she wants is the first: does pretraining scale still dominate?

She frames it as uncertainty, not risk. Risk you can put a probability on. Uncertainty you cannot. Capabilities are emergent. No one knows how algorithmic efficiency will offset more scale, what new data-generation tricks will do, or how much oil is left in test-time scaling. It is very hard to predict. What she cannot picture is a large lab trying to build AGI that would refuse the biggest cluster it could have if the capital were free. That, more than any spreadsheet, is the tell: there is still meat on pretraining, even if the returns are likely to be less efficient than they used to be.

They give themselves a free pass for New Year's predictions in February — episode 100, and the vibe of 2025, Guo jokes, is that you can just do things. Gil's five bets: the foundation-model market at least partly consolidates, starting in image, video, and voice, with secondary LLMs following, faster if the FTC is friendlier than the last regime. New races open in physics, biology, and materials while scaling, including reasoning, continues. Vertical AI applications become the era — Harvey in legal, Decagon and Sierra in customer success, Cogent in medical scribing — and a subset grow more agentic, as Cognition already has. Self-driving takes a lot of attention: Tesla and Waymo on Full Self-Driving and robotaxis, with Applied Intuition as a dark horse in the automotive stack. Consumer experiments, from startups and incumbents, get large in a way they have not; a resurgence may take a while, but it will come. Agents and robots show a glimmer, not a fire.

Guo says they agree on more than they disagree. Agent is a fuzzy word; if it means multi-step tasks in a real environment with action, not just generated content, it is already happening in security, support, and SRE, and coding got there earlier. Copilot companies will extend into agents because they always want to take more on. The input to a broader consumer wave, for her, is much more capable small, low-latency models. She does not believe in edge compute for its own sake — nobody cared — but if local inference is transparent and free, you can ship free products, and plenty of that value will live in the browser as part of an application stack. She expects technical proof of robotics breakthroughs and generalization this year, not deployments. Reasoning, still so new it is mispriced, is as much an improvement in reliability as in the complexity of the task.

The mistake she has made as an entrepreneur and investor is to watch something fail on a technical issue and assume it will not work. In AI you have to look again and again, because things begin to work quickly. A last observation from Conviction's embed program and the broader portfolio: innovation has diffused past the tip of the spear. More scientists who know a domain are picking up the tools. In biology and materials, the bottleneck is data, and the type of data, and capturing it may be a biotech innovation rather than a computer-science one. She thinks this may be the year something interesting happens in health, where you need specialized data that is still less hard than biomolecular design. For episode 200, Gil imagines two agents teaching humans, with the hosts reduced to nodes. Guo's fork is an RLHF farm or a beach in Ibiza, post-abundance. The usable residue of the joke is simpler: treat Deep Research as a junior analyst, install a Gell-Mann alarm, pick models by cost curves, and change "this doesn't work" to "this doesn't work yet."