Meta’s $197 Billion Day Was a Partial Answer to AI’s $200 Billion Question

Sarah Guo and Elad Gil skip the guest and start with models. Google’s Gemini 1.5 is the headline: competitive performance, and a million-token context window. Startups such as Magic had already talked about 5-million-token models, Gil notes, but a frontier lab shipping a million-token product is a different kind of signal. He thinks long context will matter more than people expect in biology. Fold a protein with a short window and you often fail to encapsulate the chain. The average human protein is something like 300 amino acids; plenty of them are dramatically larger. Those sequences simply do not fit.

OpenAI’s Sora is the other demo that does not look like a demo. Gil calls the video beautiful to watch and nods at Pika and other video-model companies doing interesting work. Then Mistral, which launched Mistral Large and a chat product named Le Chat — Gil cannot resist calling it Le Big Mac. The joke sits on top of a serious point. The company went from founding to almost GPT-4 level in nine months, with small efficient models, a large model, chat, and multiple languages. Execution velocity, he says, is the impressive part.

Mistral also announced a licensing deal with Microsoft. The main models on Azure, as Gil counts them, are now OpenAI, Llama, Mistral, and some Microsoft models. A European lab that did not exist a year earlier is on that list. Guo adds that the design space still wants state-of-the-art capability, which Mistral is chasing, but the company has talked from the beginning about efficiency, latency, and serving different use cases — and about being long-term proponents of retrieval.

One live research argument is that RAG dies once context is long enough: why retrieve if you can stuff millions of tokens into the prompt? Guo is on the other side. A larger window opens a trade-off space. You can retrieve more and spend less on model reasoning, or the reverse. You can work against a specific data set instead of retraining or packing everything into context. Gil’s view is simpler: we are going to have both.

A second under-discussed shift, he says, is where the compute happens. Before LLMs, a lot of human-style reasoning work happened at inference. A better poker AI trained a model, then did tree search and extra work when it actually played. He expects agents and reasoning to look more like that: heavy inference, then feed the traces back and keep training. The current pattern is spin up a data center, spend $100 million over twelve months, produce a file, and use that file for a year while you train the next one. Continuous upgrading would break that cadence. It is early in the cycle. All of these things, he thinks, will move.

Does Gemini change the Google question? Guo says the research question is answered. Gemini is impressive; internal work on function calling and multimodality that has not shipped yet is impressive too. Google has distribution, consumer behavior, search-query value, ad auctions, a great research team, and enough GPUs. The live question is whether management can steer the company to compete, against every other demand from the employee base and every mission that is not brokering the world’s information and growing market cap.

Gil is more bullish after 1.5. A year and a half ago he wrote that Google was a sleeping giant; once it woke, progress would be fast. Mistral built a company from scratch, which is extremely hard even with open source. Google has accelerated over the last two or three months, with launches that are actually externally usable. It has compute, web and YouTube data, optional personal corpora in email and Docs, and talent. What was missing until recently was the will. Competitive pressure seems to have brought the will back. Velocity, he says, might only increase from here.

Where general LLMs are still not as capable as people want is, Guo says, every domain — and two in particular. Ask ChatGPT to design a DNA sequence that can express CRISPR-Cas9 and it cannot. Cell design, protein design, protein optimization, drug discovery, and target identification are places where researchers are already getting striking results from transformers and diffusion models. That is a different kind of reasoning and a different data set, with the context-window constraints Gil mentioned. Robotics is earlier, partly because of data, though there now seem to be reasonable ways to generate some of it. Gil would not be surprised if 2024 and 2025 became years of model proliferation: chemistry, materials, physics, math, biology, robots, each growing its own models.

The other model-side wave is agents, and some of the new companies are not just stacking a bigger LLM. They are looking back at AlphaGo, poker, Diplomacy: sequential action on changing information, self-play for a billion games, heuristics and algorithms at inference rather than only at training. That body of knowledge, Gil thinks, is about to show up as products. It will take six to twelve months. The wave is coming. Guo’s version of the same news is more operational. People got smarter about agents as parts of systems. The 2023 picture was an agent in a general environment, compounding failure across ten tasks. The 2024 picture is an agent inside a loop that can reinforce it — a game, a constrained web app, or a domain such as code generation where you can sample and validate. In Q3 and Q4 of last year you could still honestly ask whether any of this was useful. Now the answer is that it is, in the right box.

Early products went too broad too early: a browser plugin that watches everything you do and then does everything you do. The targeted version is better RPA, better code, better customer support. Some teams are even buying human-feedback data to post-train, which tells them the distribution and the scale they might have to pay for. That, Guo says, turns an open-ended, untenable agent problem into a scoped one: how much will it cost to make a particular task work. They will still get to the general case. Rapid success is on the narrow front.

Guo’s line on earnings is dry: earning money is an excellent idea, and Jensen understands this better than everybody else. The worry had been that GPU demand was a short-term scramble, and that growth would calm once supply caught up. Jensen’s comment that NVIDIA expects to stay supply-constrained through the year killed that story. Demand, she says, is much larger than most people expected. The upgrade cycle underneath is A100 to H100 to H200 to B100, with a training-efficiency incentive at every step. One of her portfolio companies is buying in the tens of thousands of GPUs and skipping ahead to B100s. They described it as free money on training efficiency. It is a funny phrase, she notes, when you are spending hundreds of millions of dollars. If others feel the same way, the server market keeps expanding.

Startup rounds get the headlines. Gil says the under-discussed fact is that most of the spend is hyperscalers and other clouds building out. Enterprise adoption is still very early, and even so Microsoft said Azure revenue grew 5 percent from AI-related products. If Azure is about $25 billion a quarter, that is $1 to $1.5 billion a quarter of new AI spend, $5 or $6 billion annualized, with two or three years of real enterprise adoption still ahead. The giant LLM rounds are not venture checks. They are Amazon, Google, Microsoft, Salesforce, and NVIDIA. ServiceNow already had a blowout quarter, partly due to AI. The replacement cycle and the big-tech spend are the same story from two sides.

AI years, Guo says, are dog years, so a month back is ancient. David Cahn at Sequoia wrote "AI’s $200 billion question": given the spend, you need on the order of hundreds of billions in return to justify it. Meta’s earnings beat, for her, was a one-day answer. The company said it would spend $30 to $37 billion on 2024 capex, driven by AI and servers. Mark Zuckerberg talked about 600,000 H100-equivalent units and "no room for other people." Years of capex that made shareholders angry showed up as better targeting, conversion, recommendations, engagement, and ad ROI, plus the cost controls the rest of the industry is running. The session added $197 billion of market cap, the biggest single-session add before NVIDIA’s own beat. $197 billion of enterprise value on $25 to $30 billion of capex: you should keep doing it.

Yuri Milner once showed Gil a chart of aggregate startup market-cap growth versus the FANG cohort of the time. The top companies added five or ten times the market cap of the entire startup ecosystem. He thinks the short run of AI may rhyme: monster AI companies will eventually emerge from scratch, but for the next few years a lot of the incremental market cap will sit with existing firms, plus a few model labs. That is one reason both of them keep joking that it is a good time to run a hedge fund. Guo would actually run one on a single lens — long and short companies by how AI hits the business. Plenty of services firms sit in the sights of automation. The only question is which management teams have the capital, talent, guts, and conviction to invest the way Zuckerberg did while people were furious. Her guess, especially for services firms, is that most will not make the transition.

Klarna posted the numbers that make the abstract concrete. An OpenAI-powered assistant, in its first four weeks, handled 2.3 million customer-service chats — two-thirds of inquiries. Satisfaction was on par with humans, accuracy was higher, repeat queries fell 25 percent, and errands resolved in two minutes versus eleven. It is live 24/7 in more than 23 markets, in more than 35 languages, and it did the equivalent work of 700 full-time agents. Klarna has about 3,000 full-time agents, so the bot cut the need by about a quarter. The company also flagged the longer social question. Gil’s version is narrower and sharper: how many years before every enterprise realizes it can cut certain kinds of support with simple apps? Adoption, Guo thinks, will go use case by use case. Early adopters take the risk. Once one payments company shows a better customer experience or a real cut in operating cost, the rest of the sector has to follow to stay competitive. This stuff happens slowly and then all at once. They think we are still in the slowly phase.

Gil had his team look at global services. U.S. software spend is about half a trillion dollars a year. Payroll for human services that generative AI can probably touch is $3.5 to $5 trillion. Convert just 10 percent of that into AI revenue and you have recreated the entire U.S. software industry in market cap. Vertical by vertical you can ask where the money goes: internal tools, a new company, or an incumbent such as Zendesk. Klarna built in-house. You can also imagine a Sierra, or whoever, owning the segment for companies that will not build. Feedback loops are cheap here: thumbs-up, thumbs-down at the end of a session, a training set that should make the product better the more it is used. Guo looks at who already did it. Meta, Square, Shopify, Klarna — anyone with a large commerce or SMB-advertiser base and a technical team — have all shipped something, and it is working. If it works in payments and commerce, everybody else should follow. Retailers who can stand up e-commerce may still not build support bots; that is where an incumbent or a new company takes the segment.

Gil keeps an internal list of companies that should exist. Most of the obvious ones do not, or barely do. He splits AI talent into three waves. First came researchers, closest to the models, who built the early model companies and early apps — Perplexity, Harvey, people who already knew. Second came infrastructure people. Third will be application builders, many of whom did not notice this was important until ChatGPT, about fifteen months before this taping. Nine months to leave a job, a few months to find a co-founder, a few months to prototype: the app wave has barely shown up. A half-dozen medical-scribing apps are growing fast. Whole other spaces have nobody. Guo repeats the in-joke: foundation-model companies are here to replace all the jobs, and they do not understand any of the jobs. The opportunity is a good engineer who actually knows a domain, experimenting on the fringe of that domain. Her team’s line at an AI-in-production conference: 2023 was infrastructure, 2024 is when applications begin to appear.

If someone is already making money, someone else wants a second source. NVIDIA’s claimed defensibility, as Gil lists it, is chip performance, CUDA, and interconnect. The interconnect piece arrived with Mellanox, an Israeli company NVIDIA bought in 2019 for about $5 billion, large relative to NVIDIA’s market cap at the time. CUDA has been in the works for years. Minimally a challenger needs performant silicon, a way to use it, and a way to scale it. AMD is the incumbent chase. On the startup side he names Groq, with very fast inference for open-source and language models, and Cerebras, which took a different approach to the chip itself. Newer names such as Etched have talked about chips specialized for transformer architectures. Historically the most advanced chip markets have a leader and a second place: Intel then AMD in microprocessors; Qualcomm and Arm in mobile, now at $176 billion and $140 billion, closer than they used to be. He is not sure how that pattern holds in AI.

Jensen talks less about transistors on a die and more about memory management and full-system performance — what someone has called Jensen’s law. Guo adds manufacturing. Even a fabless designer still has to get capacity at TSMC, climb yield, and hit a price. CUDA has been around since 2006. Two trillion dollars of market cap and more demand than NVIDIA can fill make the economic pressure to challenge it higher than ever. The moat, she thinks, is still really deep. A latency-optimized system is one path. A transformer-only chip is another, and that is a bet on architectural stability. For a long time that bet looked good. For the first time in a while there is interest in state-space models, with companies such as Cartesia, which means the architecture itself has a little give. Giants with one internal workload — Meta, Google, ads, search, recommendations, spam — do not need a full software ecosystem. They need one application to run well, which is how you get TPUs and Trainium. She would love to meet a company in this area that gets her over the edge. She has not, even in a market this fertile.

The fab conversation is TSMC and a short list of others that can handle volume. The U.S. push to bring manufacturing home has hit environmental reviews and process delays on plants that take years to build. Japan, Gil says, is actually putting up fabs and may emerge as a geopolitical second source, a hedge on Taiwan. The TSMC CEO talks as much about human capital and culture as about capex. Guo wonders whether the answer is Japan, Mexico, Southeast Asia, a broader supply chain — or robots making chips. Gil pushes back on the idea that America cannot do this. Intel has manufactured in the U.S. for decades, high yield; Texas Instruments did too. Capability exists. Culture is concentrated in Taiwan and, secondarily, Korea, but the first problem is breaking ground. Intel’s U.S. fabs have been fine and also behind on process. That, they agree, looks like an Intel problem more than a national-talent problem. The question is yield and cost, and maybe for some period you do not need the chips to be fully cost-competitive.

Gil’s honest summary of the field: the more he learns about AI, the less he knows, which is the opposite of every other domain he has worked in. Knowledge usually compounds. Here a week of news can rebuild the world model. He calls it a fun exhaustion, the early slope of an exponent. A lot of work was held back inside large companies; it is flourishing outside; that creates competitive pressure; big tech reacts; people leave to start companies; big tech funds those companies at huge late-stage rounds and funds most of the industry’s compute — an order of magnitude, maybe two, more than venture. Startups accelerate big tech, which spawns more startups, which get funded by big tech. Guo would add a commercial loop: once a Klarna-style tool actually moves a P&L, the rest of the competitor set has to do it. Narrative drift — CEOs saying AI because the market wanted the word — is already showing up as budget. An investment-bank survey she cites has Fortune 1000 IT budgets going to 5 to 8 percent this year instead of 3 to 5 percent, all because of AI. That is roughly two times. They close on a joke name for the product loop: RLPA, reinforcement learning through product-adoption feedback.