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, I'm your boy Tony DeLuca, and welcome back to Barely Possible. We've got a leaner menu today than some of these monster Anthropic-IPO weeks, but there's a real piece of steak in here if you know where to look, so grab your coffee and let's dig in.
Here's the thing about today. A lot of what crossed my desk is old news dressed up like fresh news. Mistral's whole back catalog from last year showed up at my door, Meta engineering posts from last fall, a Google blog dump from their May conference. None of that is new just because somebody collected it this week. So I'm going to be honest with you about what's actually current and what's just been resurfaced, and we're going to spend our real time on the one thing today that genuinely tells you something about where the industry is heading.
And that thing is Microsoft. Specifically, a single observation from Sean Wang, who you might know online as swyx, posted yesterday, June second. He wrote, and I'll read it straight: "you have to give Microsoft AI props for training all these in-house from scratch and getting ALL of them to near-SOTA. Mustafa built a full fledged neolab inside Microsoft in 2 years, that now MS fully controls from chip to model to harness. Absurdly impressive."
Now hold onto that phrase, "from chip to model to harness," because that's the whole ballgame, and we're going to come back and take that apart properly in a few minutes. But first let me set the table with the stuff that's lighter weight, so when we get to the main course you understand why it matters.
Let's start with the only other genuinely current item on the board, which is from Anthropic, posted June second. They put out a note about expanding something they call Project Glasswing. Now I want to be straight with you about what I actually have here, because the headline is the whole thing as far as substance goes. Anthropic is expanding Project Glasswing. That's the verifiable claim. I'm not going to sit here and invent a bunch of detail about what Glasswing is or pretend I've got the internal roadmap, because that's exactly the kind of thing where a host gets out over his skis and starts narrating a story that isn't in front of him. What I can tell you is the pattern this fits into, and the pattern is consistent.
If you've been listening this week, you know Anthropic has been on an absolute tear. We covered the confidential S-1 filing with the SEC yesterday, sitting on top of that $65 billion Series H at just under a trillion dollars in valuation. Before that, the Milan office, the Seoul office, the KPMG deal putting Claude in front of 276,000 employees. So when this same company quietly says it's expanding a named project, the reasonable read is that this is one more piece of a company in full land-grab mode, building out programs and partnerships in every direction while it preps to go public. I'm flagging it so you've heard the name. If Glasswing turns into something with real teeth, we'll come back and give it the full treatment when there's meat on the bone. For now, it's a marker on the map, not a destination.
Now let me clear the decks on the resurfaced material, because there's a lot of it and I don't want anybody thinking these are today's announcements. A whole pile of Mistral news washed up this week, but none of it is new. We're talking about Le Chat's relaunch from February of last year, Le Chat Enterprise from last May, the Mistral Agents API, Mistral Code, their Deep Research preview, the MCP connectors and Memories work from last September, Mistral AI Studio from last October, and Mistral OCR 3 from this past December. That's basically Mistral's entire 2025 product calendar showing up in one batch.
And look, I'm not going to walk you through each one like it's a fresh launch, because it isn't. But there's a story in the aggregate, and it's worth ten seconds of your attention. Lay those announcements end to end and you're looking at a company that, over about a year, went from a model lab to a full product stack: a consumer assistant with apps on both phones, enterprise tiers, an agents API, a coding tool, a deep research feature, memory, enterprise connectors over MCP, a production platform, and document OCR. That is the European answer to the same playbook everybody's running. Every serious lab is trying to own the whole vertical now, not just ship weights. Hold that thought too, because it rhymes with the Microsoft story we're building toward.
Similar deal with the Meta engineering posts. There's a piece on Meta's infrastructure evolution from last September, a deep technical write-up on scaling LLM inference with tensor, context, and expert parallelism from last October, and an open-source release called RCCLX for GPU communications on AMD platforms from this past February. All resurfaced, none of it from this week. I'm not going to take you down into the kernel weeds on parallelism strategies, that's not what most of you need and it's not what this show is for. But the RCCLX one is worth one sentence as a signal: Meta open-sourcing tooling specifically to make AMD GPUs talk to each other better is a small vote against the idea that Nvidia owns the data center forever. When the biggest buyers start investing engineering hours into the alternatives, that tells you something about where they want their leverage. File it.
And then there's the Google blog dump from their May conference. A hundred things announced at I/O, the Dialogues stage recap with Pichai, twelve keynote moments, the Beam group meetings experiment, community investments in Missouri, the Waterloo Futures Lab prototypes, and a quiz they vibe-coded in Google AI Studio. Recent reports, a couple weeks old now, not today's news. We covered the substantive piece, the AI Mode search changes, in an earlier episode, so I'm not going to relitigate the conference. The one thread I'll pull, because it connects to our main story, is that little detail about Google vibe-coding their own promotional quiz inside their own AI Studio. It's marketing, sure. But it's also Google demonstrating the same instinct everybody's got right now: use your own model, in your own tooling, to build your own stuff. Vertical integration as a lifestyle. Which brings us right back around to Microsoft.
So let's dig into the one that actually matters today: Microsoft quietly building a full-stack AI lab inside itself, and what swyx is really telling you when he says chip to model to harness.
Let me read the observation again so we're working from the exact words. Sean Wang wrote: "you have to give Microsoft AI props for training all these in-house from scratch and getting ALL of them to near-SOTA. Mustafa built a full fledged neolab inside Microsoft in 2 years, that now MS fully controls from chip to model to harness. Absurdly impressive."
Now, who's Mustafa? That's Mustafa Suleyman, who runs Microsoft AI. And the thing swyx is reacting to is that Microsoft has, by his read, trained a whole family of models in-house, from scratch, and gotten all of them to near state-of-the-art. Not one flagship they can point to in a press release. A spread. And the part that made him sit up is that phrase about control: chip to model to harness.
Let me unpack what each of those three words means for you as a builder, because this is the part that actually changes how you should think about your own dependencies.
Start with the chip. For years, the entire AI industry has been renting its existence from Nvidia. If you wanted to train a serious model, you stood in line for H100s and later Blackwells, you paid whatever the market said, and your timeline was hostage to allocation. Owning the chip layer, or at least having your own silicon in the mix, means you're not standing in that line on the same terms as everybody else. Microsoft has been building its own accelerators. So when swyx says they control from the chip, he's saying Microsoft can, at least in part, make the metal their models run on. That's the bottom of the stack.
Then the model. This is the one that's genuinely surprising, and here's why. For the last couple of years, Microsoft's AI story was basically: we own a giant chunk of OpenAI, we put GPT into everything, we're the distribution arm. The models were somebody else's. The narrative was that Microsoft didn't need to train frontier models because it had the OpenAI partnership locked up. And what swyx is flagging is that, quietly, over about two years, Suleyman built a real lab inside Microsoft that trains its own models from scratch and gets them near the frontier. That's a strategic hedge of enormous size. It means Microsoft is no longer fully dependent on OpenAI for the brains. If the partnership ever frays, if pricing fights get ugly, if OpenAI decides it wants to compete more directly in Microsoft's enterprise turf, Microsoft now has its own models to fall back on. That's not a side project. That's an insurance policy worth tens of billions.
And then the harness. Now, longtime listeners know I have feelings about the word harness, because it got so overused this spring that I almost banned it from the show. But here it means something specific and important. The harness is the layer that sits around the model and actually makes it do useful work: the loop that lets it keep going, the tools it can call, the way it manages context, the scaffolding that turns a raw model into a working agent. Owning the harness means you control the experience the customer actually touches. The model is the engine, the harness is the car. And swyx is pointing out that Microsoft now owns all three layers: the metal, the engine, and the car.
So why does that matter to you, sitting there building a product? Three reasons.
First, it tells you the moat conversation has moved. For a while everybody argued about whether the model was the moat. Then people said no, it's the harness, it's the scaffolding. The smarter read, and the one this Microsoft story supports, is that the real position of strength is owning enough of the stack that no single supplier can squeeze you. Microsoft isn't betting on being the best at any one layer. It's betting on controlling all of them so that its margins and its roadmap aren't hostage to anybody. If you're a founder, the lesson isn't go build your own chip, obviously. The lesson is: know exactly where your single points of failure are. If your entire company sits on one model provider's API at full price, and that provider can change your unit economics with a pricing email, you don't have a company, you have a tenancy.
Second, it reframes what near-state-of-the-art actually buys you. Notice swyx didn't say Microsoft built the best model in the world. He said they got all of them to near-SOTA. And here's the quiet truth of this era: near-SOTA that you fully own and control is often worth more to a big platform than best-in-class that you rent. If your model is ninety-five percent as good but it runs on your chips, in your harness, at your cost structure, inside your distribution, you win on economics even while you lose on the leaderboard. That's a deeply uncomfortable idea for the labs that have been competing on benchmark points. The benchmark crown is worth less than it used to be when the buyer can get most of the capability and all of the control by going in-house.
Third, and this is the one I want you to really chew on, it tells you where the big players think the constraint is going to be. They are not building this much vertical control because compute is cheap and abundant. They're building it because they expect it to be scarce, contested, and expensive. When Microsoft invests two years and untold billions into owning its own silicon and its own models, that's a company planning for a world where you cannot just buy your way out of a compute shortage on the open market. Owning the stack is what you do when you think the supply is going to get tight. And you've heard versions of this from every direction lately, the whole token-cost reckoning hitting corporate America, companies burning through annual AI budgets in a few months. The big platforms aren't waiting around to see how that shakes out. They're vertically integrating now so they're not the ones left holding an empty cup when the well runs low.
Let me connect this to the resurfaced stuff I rushed past earlier, because now it should click into place. That whole Mistral product catalog I listed? Same instinct, smaller scale: own the model, own the agents API, own the coding tool, own the production platform, own the document pipeline. That Meta RCCLX release making AMD chips work better? Same instinct: reduce dependence on a single chip vendor. Google vibe-coding their own materials in their own studio with their own model? Same instinct. Everybody, from the giants to the European challenger, is reaching for control of more of the stack. The Microsoft story is just the cleanest, most complete example of the pattern, and swyx happened to put words to it yesterday.
Now, let me put on the skeptic hat for a second, because that's my job and you'd be right to want it. "Near-SOTA across the board, fully controlled from chip to model to harness" is one builder's enthusiastic read of a marketing-adjacent moment, not an audited fact sheet. I don't have independent benchmarks in front of me proving every one of these Microsoft models lands where swyx says it does. "Near" is doing a lot of work in that sentence, and "from scratch" can mean a lot of things depending on what you count as scratch. So take the specific superlatives with the appropriate grain of salt. What I'm confident about is the strategic direction, because the direction is corroborated by everything else on the board today and everything we've been tracking for weeks. The vector is real even if you want to argue about the exact magnitude.
And here's the practical takeaway for the person actually shipping product. You are not Microsoft. You're not going to own your own chips. But the principle scales down. Ask yourself the chip-to-model-to-harness question about your own stack. Where's your chip-level dependency, meaning your raw compute and where it comes from? Where's your model-level dependency, meaning whose brains are you renting and on what terms? And where's your harness, meaning how much of the actual product experience do you own versus how much is somebody else's scaffolding that could change under you? The companies that survive the squeeze coming this year aren't the ones with the cleverest prompts. They're the ones who looked at that three-layer stack and made sure they weren't fully exposed on any single layer to a vendor who could change the rules on a Tuesday.
That's the deep dive, and honestly that's the spine of the whole episode, so let me tie up the loose ends and we'll get out of here.
One more current item I want to mention, partly because it made me smile and partly because there's a real point under it. swyx had another post from June first that's basically inside-baseball banter. He's riffing with some folks online, mentions he's never heard a particular company called a, quote, "kwa" startup, and says now he just wants somebody to make him a "bak kwa" model. For those who don't know, bak kwa is a sweet, grilled dried meat, a barbecue jerky you find around Lunar New Year. It's a joke, a pun, nothing more. I'm not going to manufacture a trend out of a guy being playful on the timeline.
But I'll say this, because it's the kind of thing that's easy to miss: the fact that the people building and analyzing this industry are loose enough to make food puns at each other tells you something about the culture of where the real conversations happen. The signal isn't the pun. The signal is that if you want to actually understand where this stuff is going, the marketing blogs and the conference recaps are the lagging indicator. The leading indicator is what the practitioners are saying offhandedly to each other in public, the way swyx put words to the Microsoft full-stack thing in two sentences before any glossy think-piece will. Pay attention to the offhand observations from the people with their hands on the actual machinery. That's where the real read lives.
Let me give you the throughline before I sign off, because I want you to leave with one idea, not seven. The story today, underneath the Microsoft post and the resurfaced Mistral catalog and the Meta chip tooling and even Anthropic quietly expanding its named projects while it preps to go public, is the same story. Everybody who's serious is racing to control more of their own stack, because they all expect the era of cheap, abundant, rentable AI to end, and they don't want to be the ones standing in somebody else's line when it does. Microsoft just gave us the clearest picture of what that finished move looks like: chip to model to harness, all in-house, all controlled. The question for you isn't whether you can match that. You can't. The question is whether you know where you're exposed, and whether you've got a fallback for the day your one critical supplier decides to change the deal.
That's our menu for today. Lighter portion, but I'd rather feed you one good thing than pad it out with reheated leftovers and pretend they're fresh. As always, the links are in the show notes if you want to go read the originals yourself, and you should, because you should never just take my word for it. I'm Tony DeLuca, this has been Barely Possible, and I'll be right here in your ears tomorrow. Be good to each other out there.