Last Week Ignite — 5.10.26
Past the Model
I spent a couple of hours over the weekend reading DeepMind’s AlphaEvolve impact post from Tuesday. The number that stayed with me came from Schrödinger, the computational chemistry company. They said AlphaEvolve had made their machine-learned force-field training roughly four times faster. Force fields are the physics simulations chemists use to model how molecules behave around each other. They sit upstream of much of early-stage drug discovery. Four times faster means a big chunk of that pipeline got materially cheaper this week, because an AI agent rewrote how the underlying tools were built.
That’s a narrow technical event. It’s also the cleanest signal I saw all week.
Last week’s letter argued that the dominant AI investing question, which model wins, had stopped being the useful one. This week the news kept proving that point in different categories. Workflow ownership in customer service. Discovery-loop acceleration in computational chemistry. Optical and power scarcity around Nvidia’s announcements with Corning and IREN. Adoption depth in Microsoft’s diffusion index. Embodiment as a structural problem in robotics from Genesis AI. Each of those is a different bottleneck. None of them is a model.
Sierra raised $950 million on Monday at a valuation north of $15 billion. The company runs agents inside customer-service operations at more than 40% of the Fortune 50, and says those agents are handling billions of interactions. Two years ago the support-chatbot category was a graveyard of wrappers. The market this week priced one company in it as durable infrastructure.
The thing I keep coming back to with that round is what wasn’t in the pitch. Sierra didn’t win a $15 billion mark by training a better model. It won by integrating into systems of record, by building evaluation loops customers will trust, and by getting to enough volume that switching costs are now real. The model is a component in that stack. The stack is the product.
Read that next to the AlphaEvolve update and the picture sharpens. AlphaEvolve says a model can now meaningfully accelerate the tools that drive future discovery. Sierra says a software company can lock down a category if it owns enough of the operational surface around the model. Different layers, same pattern. The investable artifact is whatever the model is feeding into and pulling from, more than the model itself.
The infrastructure side of the week made the same point in concrete. On Tuesday, Corning announced a partnership with Nvidia that will expand its US optical-connectivity manufacturing capacity by ten times, open three new plants, and add more than three thousand jobs. A day later, Nvidia and IREN said they intended to support the deployment of up to five gigawatts of AI infrastructure together. Optics are the wires that move data inside data centers. Five gigawatts is the power output of about five large nuclear reactors.
I wrote about Meta and Entergy’s Louisiana grid deal last week as a sign that generation, not GPU count, had become the binding constraint on AI buildout. This week the binding constraint got broader. The chokepoint isn’t only generation. It’s interconnects, optics, siting, financing, and the industrial supply chain that surrounds a GPU. Founders who treat compute as a utility you order on demand are about to learn it isn’t, at least not at the scale serious AI products run. The platform companies have figured this out and are building factories in response.
Microsoft’s diffusion index, published Thursday, sat next to all of this in a useful way. Global AI usage among the working-age population went from 16.3% to 17.8% in the first quarter. US usage hit 31.3%. The gap between Global North and Global South widened to 27.5% versus 15.4%. The headline I keep hearing from operators is that capability has stopped being the rate-limiting factor. The gap that matters is how fast an organization absorbs what already exists. Some companies are running agentic workflows inside their core operations. Others still have a “ChatGPT for work” pilot stuck in marketing from a year ago. Both of those things are happening at the same time, in companies of similar size, in similar industries. Microsoft just put a number on the gap.
That widening gap is going to do interesting things to enterprise software categories that haven’t fully formed yet. Adoption infrastructure, agent governance, observability, change-management software for AI rollouts. Last cycle these looked like consulting garnish. This cycle they look more like real software categories. Microsoft’s data is the first piece of evidence I’ve seen that procurement teams might start asking about absorption rate the way they currently ask about security posture.
Genesis AI gave the week its quietest but most honest piece of news. On Tuesday they unveiled a foundation model called GENE-26.5 alongside a demo of robotic hands they had designed with a partner called Wuji Tech. Management said they had decided to go full-stack because controlling hardware mattered for progress. Robotics has spent the last two years borrowing the foundation-model playbook. Train a giant model on internet-scale video, get out a generalist policy, ship it. The Genesis announcement is the loudest acknowledgment yet that the playbook has limits. Embodied systems need data nobody has, control loops that can fail safely, and hardware that physically supports the policies the model wants to run. The robotics companies most likely to compound own data, hardware, and weights as one stack. Companies that own only weights are racing toward a margin profile that won’t justify their valuations.
Kalshi raised $1 billion on Wednesday at a $22 billion valuation, doubling its mark in five months. Prediction markets occupy strange ground. They have regulatory positioning the rest of consumer fintech can’t replicate easily, and attention-capture properties closer to sports books than to investment products. The mark isn’t proof of durability. It’s proof that capital is willing to chase a clear category leader at high prices right now.
That’s the harder pattern of the week. Sierra and Kalshi got marked aggressively. The middle of the venture market did not. Founders selling into AI categories without obvious distribution moats are still telling me their seed extensions are taking three times longer than their original seeds did. Capital is concentrating into the leaders. The middle pays for it in time and dilution.
Macro confirmed the same shape. The Bureau of Labor Statistics said job openings were 6.9 million in March, hires rose to 5.6 million, first-quarter productivity was up 0.8%, unit labor costs were up 2.3%, and April nonfarm payrolls came in at 115,000 with unemployment unchanged at 4.3%. None of that is a green light. None of it is a red flag. It’s a functioning economy with enough labor scarcity to keep the labor-substitution story alive and enough budget room for buyers to keep spending if you can show them a number.
Five separate moves. One pattern. The AI question fragmented this week into a set of more specific questions, and they’re harder to answer than “which model wins.”
The questions experienced founders are sitting with right now look concrete:
What workflow do you sit inside often enough to build a proprietary feedback loop?
If model prices fall another order of magnitude, what happens to your gross margin?
What labor line item or revenue leak do you remove that a customer can quantify?
What do you own when compute gets gated by power and optics rather than GPU supply?
In robotics, what part of the data-and-control stack do you genuinely control?
The questions experienced investors are sitting with are different:
Are the marks for category leaders signal or vanity?
Where in the chain around the GPU is rent accumulating that hasn’t been priced in?
Which adoption-and-governance tools become real software categories now that diffusion data is starting to make consulting look uncomfortable?
Which late-stage names look like Sierra’s pattern, which look like Kalshi’s, and which look like neither?
What I’m watching over the next few weeks. Whether Sierra-style workflow valuations spread beyond Fortune 50 customer profiles or stay concentrated in the top tier. Whether the Corning and IREN announcements turn into permits and operating capacity rather than press-release commitments. Whether enterprise procurement starts citing Microsoft’s diffusion data the way it cites security frameworks today. Whether Genesis AI’s full-stack admission gets followed by other robotics companies quietly walking back their pure-software claims.
The week didn’t deliver one clean takeaway. It delivered a set of cleaner questions. That’s usually what real shifts feel like before they show up in returns.