Subscribe
Share
Share
Embed
A daily briefing on the AI systems, products, companies, and policy shifts that are just becoming possible.
Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_feed&feed_source=rss
Okay kiddos, pull up a chair, it's your boy Tony DeLuca and you've found Barely Possible. Today we've got a menu that's mostly one company doing the same thing the rest of the industry's been doing, except louder and with a banker on the line. We're gonna talk about a frontier lab quietly lining up to go public, a research dump that tells you what AI is actually being used for behind the glossy keynote, and a laptop chip pairing that's taking a swing at a six-year monopoly. So grab the coffee, mind the time, and let's have at it.
Let me set the table honestly, because I respect your time too much to pretend otherwise. A big chunk of what landed in front of me today is older news that got freshly collected, not freshly made. Mistral's got a whole row of product announcements stretching back to early last year — Le Chat, the enterprise tier, the agents API, OCR — all of that is months old, some of it more than a year old. Same with a stack of Meta engineering posts on vector search, concrete chemistry, and inference parallelism. Good stuff, but it's a catalog, not a headline. I'm not gonna stand here and tell you a January 2025 cat-pun press release is breaking news. So I'll thread the genuinely new material through, flag the old stuff as background where it actually helps, and skip the rest. That's the deal.
Let's start with the one that actually matters most for anybody building a company right now. Anthropic confidentially submitted a draft S-1 to the SEC. That's the current, dated-today piece, and it's the kind of thing you don't bury.
Now, what does that actually mean in plain English? A confidential draft S-1 is the paperwork a company files with the SEC when it's getting ready to go public, but it files it privately first. You don't see the numbers yet. The public doesn't get the prospectus. It's a way to start the formal IPO conversation with regulators, get feedback, clean up the filing, all without tipping your full financial hand to competitors and the press before you're ready. Companies do this when they're serious but not yet committed to a date. It is a starting gun, not a finish line. Anthropic could pull this, delay it, or sit on it for a long stretch. But you don't file an S-1, even confidentially, on a whim. Lawyers and bankers get paid a lot of money to make that happen.
Here's why this lands the way it does. We talked a couple days back about Anthropic's Series H — they closed sixty-five billion dollars at a nine hundred sixty-five billion post-money valuation. That's the number from late May, and I'm not gonna re-chew it because we already did. But put the two together and the picture sharpens. You raise sixty-five billion in private money, you're sitting just shy of a trillion-dollar valuation on paper, and within days you're quietly handing the SEC a draft S-1. That's a company that has, more or less, run out of bigger private rooms to raise in. When you're approaching a trillion dollars, the pool of private investors who can write meaningful checks gets thin. The public markets are where the really deep capital lives. So this is the logical next move for a business that needs to keep feeding an extraordinarily hungry compute habit.
And that's the part I want you, the builder, to sit with. Why does a frontier AI lab need public-market money? Because the cost of staying at the frontier is brutal and it never stops. Training runs, inference at scale, data centers, the chips, the power. We've covered the token-cost squeeze on this show more than once — the enterprise side where the logs look busy but the bill is what hurts. This is the supply side of that same equation. The model maker is hunting for capital on a scale that private venture, even at these absurd levels, can't comfortably sustain forever.
So what should you watch next? Three things. First, whether and when the actual numbers go public, because a confidential draft eventually has to become a public prospectus if they go through with it, and that's when we finally see real revenue, real margins, real burn. Anthropic's been described as running at something like a forty-seven billion dollar run-rate, but a run-rate is a snapshot, not a P&L. The S-1, when it's public, is where the costs come out of the shadows. Second, watch what this signals to OpenAI and the rest. When one frontier lab moves toward the public markets, the others have to think hard about their own capital path. It's a tell. Third, and this is the founder-specific angle: an IPO changes how a company behaves. Public companies answer to quarterly expectations. They get more conservative about pricing experiments, more careful about giving things away, more disciplined about which bets they fund. If you're building on top of Anthropic's API, a path to the public markets means the era of generous, growth-at-all-costs pricing has a clock on it. The free lunch gets a menu, and the menu gets a price.
That's the through-line I keep coming back to. The economics of staying at the frontier are now so large they're reshaping the corporate structure of these labs. A confidential S-1 isn't a product. It's a confession about how much this all costs.
Now, while we're on Anthropic, let me clear the decks on the other items from them, because there's a whole row of recent reports and resurfaced posts and I don't want to pretend each one is a fresh bombshell. There's the two-hundred-million-dollar partnership with the Gates Foundation, the acquisition of a company called Stainless, both of which are older items that surfaced in the feed but aren't from today. There's a recent report that KPMG — the big accounting and consulting firm, a workforce north of two hundred seventy-six thousand people — is integrating Claude across its core business in a strategic alliance. There's the Milan office, a Korea office on the way with a new representative director named KiYoung Choi appointed ahead of the Seoul opening. And a piece on widening the conversation on frontier AI.
I'm not gonna deep-dive any of those individually, but notice the pattern, because the pattern is the story. Enterprise alliance with a global consulting giant. Offices opening in Italy, in Korea. A draft S-1. This is what a company looks like when it's growing up in a hurry — landing the white-shoe enterprise accounts, planting flags in international markets, and getting its house in order for the public markets. The KPMG deal in particular is worth flagging for builders, because when Claude goes into a firm with a quarter-million employees, that's not a pilot. That's distribution at a scale that changes the default. Big consulting shops are how AI gets pushed into thousands of mid-sized companies who'd never integrate a model themselves. KPMG becomes a channel. That's the quiet enterprise reality underneath the IPO headline.
Now let's shift from the cap table to the keyboard, because there's a recent report out of Google that actually tells you something useful about how regular people are using this stuff. The piece, How AI Mode is changing the way people search in the U.S., from Shivani Mohan, is a recent report — a couple weeks old — on AI Mode in Google Search.
Now, I'll be straight with you: what I've got in front of me is the framing, not a giant table of numbers, so I'm not gonna invent statistics. But the angle itself is worth your attention if you build anything that depends on search traffic, which is to say, almost everybody. AI Mode is Google's push to answer your query directly, conversationally, instead of just handing you ten blue links. And the whole reason Google's publishing a piece on how it's changing search behavior is because the behavior is, in fact, changing. People are asking longer, more complicated questions. They're treating the search box less like a card catalog and more like a person they can interrogate.
Here's the builder's takeaway, and it's not a happy one if your business model is getting found. When the search engine answers the question itself, the click to your website doesn't happen. The user gets what they came for inside Google's box. For a decade, the entire content economy was built on the assumption that you write something good, Google sends you traffic, you monetize the traffic. AI Mode chips away at the middle of that chain. If you're a founder whose customer acquisition leans on organic search, you should be treating this as a slow-motion structural shift, not a feature update. The question isn't whether AI Mode is cool. It's where your traffic comes from when the answer never leaves Google.
And this connects to the bigger Google story sitting in the feed, which is the whole I/O 2026 pile. There's the hundred-things-we-announced recap, the Dialogues stage with Sundar Pichai, twelve major moments, nine demos of Gemini Omni and Gemini 3.5, a quiz they vibe-coded in AI Studio, and a current piece, How we used Gemini to build Google I/O 2026, from Marvin Chow. These are mostly recent reports — the event already happened, most of this is the after-party recap content.
I'm not gonna march through a hundred announcements. Nobody needs that and you'd hate me for it. But let me pull the one thread that matters. The newest of these, the how-we-used-Gemini-to-build-I/O piece, is Google telling you it used its own AI to produce its own flagship developer event. The vibe-coded quiz, the AI-built collateral — that's the company eating its own dog food in public and making sure you know it. It's a flex, sure. But it's also a signal about where Google thinks the value is: not just the model, but the model wired into the production pipeline, doing the grunt work of building things. Gemini Omni and Gemini 3.5 are the headline models, the multimodal push, the stuff that handles audio and video and images all at once. For you, the practical question is whether Omni's multimodal capabilities are good enough to replace a stack of specialized tools you're currently paying for separately. That's the evaluation worth running. The keynote is theater. The integration into how work actually gets done is the part that touches your roadmap.
Let me connect those two Google items honestly, because they're really the same coin. AI Mode changing search is the consumer-facing edge of Gemini getting woven into everything. The search box and the developer pipeline are both places where Google is replacing a discrete action — clicking a link, hiring a coder for a quiz — with a model doing it inline. That's the strategy. Whether it's good for the open web is a separate, harder question, and the honest answer is: probably not, if you're a small publisher.
Now let's get into something with a little more hardware grease on it, because there's a current post that genuinely surprised me. Swyx flagged it — Grace plus Blackwell chips in a laptop. Microsoft and Nvidia teaming up, in his words, to take on six years of total dominance of Apple Silicon.
Let me unpack that, because it's a real shot fired. For about six years now, since Apple moved off Intel to its own M-series silicon, Apple's laptops have basically owned the premium performance-per-watt conversation. If you wanted a thin, quiet, all-day-battery machine that didn't melt when you pushed it, you bought a Mac. Windows laptops kept losing that argument. Grace and Blackwell are Nvidia's names — Grace is their CPU line, Blackwell is the current GPU architecture, the stuff powering the big AI data centers. Putting that pairing into a laptop, with Microsoft on the software side, is Nvidia saying: we're not just going to own the data center, we're coming for your lap.
Here's why a builder should care, and it's not about which logo is on your machine. It's about where local AI inference is headed. A laptop with serious Nvidia silicon inside means running real models on-device gets dramatically more viable. Right now, a lot of the privacy-sensitive, latency-sensitive, offline-capable AI work has to phone home to a cloud GPU. If you can put a Blackwell-class chip in a laptop, you change the calculus on what you can run locally — which matters for anybody building tools where sending data to someone else's server is a dealbreaker. Healthcare, legal, finance, anything regulated. The chip in the laptop is downstream of a much bigger question: how much AI moves from the cloud back to the edge.
Now, I'm going to be disciplined here, because this is one social post and a teaser, and I'm not going to spin a full spec sheet out of a tweet. We don't have the benchmarks. We don't have battery numbers. We don't have pricing. Swyx is excited, and the framing of Microsoft and Nvidia ganging up on Apple is genuinely a fun fight to watch, but the proof is in the machines, and the machines aren't in our hands. So file this under: watch closely, judge when the reviews land. If Nvidia and Microsoft can actually match Apple on the thing Apple's best at — performance without the machine sounding like a hair dryer and dying by lunch — that's a real story. If it's another hot, loud, two-hour-battery brick with a fast chip nobody can keep fed, it's a press release. The history of Windows laptops trying to dethrone the Mac is a graveyard of impressive chips wrapped in disappointing machines. I want to believe. I've also been around long enough to wait for the receipts.
Let me sit with the Apple Silicon point for one more beat, because it's instructive for founders even if you never touch hardware. Apple won that six-year stretch not by having the single fastest chip — it didn't, on raw numbers, in plenty of cases. It won by integration. The chip, the operating system, the battery management, the thermals, all designed together by one company that controlled the whole stack. Microsoft and Nvidia are two different companies trying to glue their pieces together to beat one company that owns the whole thing. That coordination problem is exactly why these challenges have failed before. It's the same lesson that shows up in software all the time: the best individual component doesn't win, the best-integrated system does. So the question isn't whether Blackwell is fast. Of course it's fast. The question is whether two giants can cooperate tightly enough to make a coherent machine. That's a much harder problem than transistors.
Now let's go to the part of the feed that's quieter but, if you actually read it, tells you what AI is doing when the keynote lights are off. Meta's engineering blog dropped a whole back-catalog into the feed — and to be clear, almost all of it is older, resurfaced material, months old, some of it from last spring. I'm not presenting any of it as today's news. But taken together, it's a useful X-ray of how a company at Meta's scale actually deploys this stuff in production, which is a very different animal from the demo reels.
Let me pull the two that matter most for builders, both of them older posts, both of them about AI doing unglamorous engineering work. The first is Diff Risk Score — DRS — an internal Meta tool that predicts the likelihood that a given code change causes a production incident. They built it on a fine-tuned Llama model. It looks at a code change and its metadata and spits out a risk score, flagging the dangerous diffs before they go out. Think about what that actually is. It's not a chatbot. It's not vibe-coding a quiz. It's a model wired into the deployment pipeline as a safety check, quietly grading every change for how likely it is to break things. That's AI as infrastructure plumbing — the boring, valuable kind that nobody tweets about.
The second is their Automated Compliance Hardening tool, which uses LLMs for mutation testing and compliance. Again, deeply unsexy on the surface. Mutation testing is where you deliberately introduce small bugs into your code to check whether your test suite actually catches them — it's a way of testing your tests. Meta's using language models to automate parts of that and to harden compliance adherence. And the reason I drag these two old posts into today's conversation is the contrast. We spend a lot of airtime on the flashy frontier — the trillion-dollar valuations, the chips in laptops, the keynote demos. But the place where AI is quietly, durably earning its keep is exactly here: code review, risk scoring, test generation, compliance. The unglamorous internal tooling.
For you, the builder, that's the more honest signal about where the durable value is. Not the model that writes you a sonnet. The model that catches the bad deploy at three in the afternoon before it pages everybody at two in the morning. If you're trying to figure out where AI actually pays for itself inside an organization, look at the boring stuff. Look at the pipeline. That's where the ROI hides, and it's the same lesson we keep circling on this show — the value isn't in the magic trick, it's in the workflow the magic trick quietly removes.
There's a couple more older Meta items worth a one-line nod, not a deep dive. There's a piece on using AI to design lower-carbon, faster-curing concrete, using Bayesian optimization — basically letting an algorithm search the giant space of possible concrete recipes to find ones that are stronger and greener and set faster. That's AI doing materials science, and it's a reminder that not all of this is software eating software; some of it is software eating chemistry. And there's the Reality Labs research on wrist-worn devices that read the electrical signals in your forearm — surface electromyography — to let you control a computer with subtle hand movements. Both interesting. Both old. Both background. I flag them so you know the breadth, not because either broke this week.
Let me do the same quick, honest pass on the Mistral pile, because there's a long row of it and I don't want you thinking I skipped it out of laziness. It's a year-plus of product history that all surfaced at once — Le Chat the assistant, the iOS and Android apps, the Pro and Team and Enterprise tiers, Le Chat Enterprise, the Agents API, Mistral Code, Deep Research, the MCP connectors and Memories, AI Studio as their production platform, OCR 3 for documents. Read as a timeline, it tells one coherent story: Mistral spent the last year-plus building out a full-stack assistant-and-platform business, the European answer to the American labs, with enterprise integrations and an agents API and a coding tool, working its way up the value chain from raw models toward products people actually pay for. But none of it is new today. It's a resurfaced catalog. So I'll leave it as context — the European frontier player is building a real product surface — and I won't dress up a 2025 announcement as a 2026 headline. You deserve better than that.
Let me pull the camera back and tie today's genuinely fresh threads together, because there is a spine here. The Anthropic S-1 says the cost of frontier AI is now so big it's bending the corporate structure of the labs toward the public markets. The Google AI Mode and Gemini story says the same compute and capability are getting woven inline into search and into the production pipeline, quietly replacing clicks and tasks. The Nvidia-Microsoft laptop says the hardware fight over where that AI runs — cloud or edge — is heating up and reaching for your lap. And the older Meta engineering posts say that underneath all the noise, the durable value is the boring internal plumbing: the risk score on the diff, the test that tests your tests.
If you're a founder, here's the one thing I'd carry out of today. Everybody's watching the frontier — the valuations, the chips, the keynote. But the frontier is also where the costs are most ruinous and the pricing is least stable. The Anthropic IPO path is a reminder that the generous, subsidized phase of this market has an expiration date. So build like the bill is coming due. Lean on the boring, durable applications where AI removes real work and the value survives a price hike. And when somebody waves a tweet at you about a laptop chip slaying a six-year giant, nod politely, and wait for the machine to show up before you reach for your wallet.
That's the menu for today. The IPO that's really a confession about cost, the search box that keeps your traffic, the laptop picking a fight it might not win, and the quiet tools doing the actual work. I'm Tony DeLuca, this has been Barely Possible, and I appreciate you spending the time with me. Be skeptical, be curious, and don't pay full price for the hype. Catch you on the next one.