The AI map is starting to fill in

After nearly three years of frenetic growth, the AI market is starting to take shape. Investors, analysts and founders are noticing, and the conversation is shifting from what AI might become to who will actually win. Every technology wave goes through this. The internet did in the late '90s, when the chaos settled into a stable structure, and that clarity was a gift to anyone building something durable. I think we might be entering that phase now. Here's how the map looks to me today:

Frontier models: powerful, expensive, and less relevant than many thought. The frontier labs keep producing astonishing models, and the leaderboard changes almost weekly. Google's Argon put it back on top; by the time you read this, someone else may have taken the lead. And yet, I am not sure it matters. Customers swap models with a config change, and developers route across several at once. Switching costs are simply too low for frontier leadership to become a durable advantage. What decides business outcomes is context, workflow, data and cost. We may see technical breakthroughs and mega-acquihires, but I am not sure we'll see dramatic or durable shifts in market share anytime soon.

Neoclouds: the bill is coming due. The neoclouds borrowed heavily to build capacity ahead of demand, and the credit markets are losing patience. CoreWeave's bonds yield close to 15%, Firmus pulled its IPO days before pricing, and Elice fell 32% on its first day of trading. This weekend Elon Musk warned that many neoclouds face "guppy-level death and consolidation." My bet is that the hyperscalers step in, through acquisitions, capacity deals, or simply outlasting everyone. The middle layer of the stack was always a scale game, and scale games favor the biggest balance sheets.

Inference compute: the new gold rush. If training was Nvidia's to lose, inference is where everyone thinks they can win. A remarkable number of software acceleration and chip companies (GPU and CPU alike) are raising capital to take a slice of Nvidia's share, betting that inference will be bigger, more diverse and more price-sensitive than training. Some will be right. Most will not. The market is real, but the crowd chasing it is a reminder that we're still partly in the frenzy.

Security: a vast frontier, but the battle lines are drawn. AI has massively expanded the attack surface: agents holding credentials, code shipping faster than anyone can review it, models wired into every system. Security has never mattered more. Yet AI-native security players are converging on a handful of powerful ideas, most of them about real-time vulnerability remediation, fixing problems as fast as they're created. There will always be room for an independent security layer, but it has never been harder for a new entrant to break in.

Open source and dev tools: more code, less value. Open source is here to stay. Agentic coding has unleashed a faster-than-ever proliferation of new approaches to almost every coding problem. But that abundance is exactly what breaks the business model: when an agent can produce a credible alternative to your library over a weekend, it's hard to charge for it. More code is being written than ever, yet I expect very little of the AI era's value to accrue to dev tools.

Value capture: outcomes and workflows, not tokens. A popular thesis holds that the winners will be whoever gets "into the token stream" and takes a cut of every call. I don't think that's wrong, but call me somewhat skeptical. Models and inference are commoditizing and getting more efficient by the month, so catching tokens for their own sake is a claim on a shrinking unit of value. The real value lies in selling outcomes (Service as a Software) or in truly owning a human workflow in an AI-native way. Where humans disappear from a workflow entirely, only outcome-based pricing will remain. Where they don't, expect a long slugfest between incumbent software vendors bolting on AI and AI-native challengers, all fighting for the same human workflow.

Enterprise adoption: real, but slow. For all the noise, the enterprise is adopting AI at an enterprise pace. Pilots take quarters, procurement takes longer, and rewiring how people actually work takes years. That gap between what the technology can do and what organizations have absorbed is the opportunity. It leaves plenty of room for specialized vendors to build companies of real consequence by doing the hard, unglamorous work of fitting AI into specific industries, workflows and compliance regimes.

Physical AI: harder than the money suggests, and bigger than ever. Capital is pouring into robotics, autonomy, defense, energy and industrial AI, and the size of those flows makes it all look easier than it is. Bringing AI into the physical world means hardware, supply chains, safety certification, long sales cycles and messy real-world data that no model can simply scale its way past. Failure rates will be high, and many well-funded companies won't make it. But the opportunity is larger than it has ever been. Intelligence can finally reach into factories, fields, grids and supply chains, the parts of the economy that dwarf the software market itself. For founders with the patience and domain depth to get it right, this is among the most exciting opportunities I have ever seen.

What remains. Put the pieces together and a pattern emerges. The layers that looked most exciting in 2023 (models, compute, developer tools) are commoditizing, consolidating, or both. The value is accruing elsewhere: in outcomes customers will pay for, in human workflows rebuilt from the ground up, and in the slow, stubborn work of getting enterprises to actually change. The opportunity for founders is narrower than it looked three years ago, but perhaps a bit clearer. The map isn't finished, but the blank spaces are finally getting smaller. For founders who know exactly where they want to build, that's the best news in years.

Gil Dibner

Could the future of software be fluid
How do we get the best of AI without losing the soul of software?

The future belongs to young missionary teams
Why it makes more sense betting on youth in the current moment

The AI-native enterprise playbook
Ten real-time observations on a rapidly evolving playing field

No more painting by numbers
It’s the end of the “SaaS playbook.

HARD TECH

Bringing drug chemistry back onshore. Warp Sciences came out of stealth this week with $215M at an $865M valuation. Lightspeed led the round, with IVP, Mayfield, Lux Capital, Premji Invest, and Menlo Ventures participating. Angular is also an investor. The founding team is unusually strong. CEO Kevin Weil ran OpenAI for Science until April, after earlier leadership roles at Instagram, Twitter, and Planet. COO DJ Patil was the first US Chief Data Scientist. CSO Sol Reisberg previously led medicinal chemistry at Terremoto Biosciences. CTO Mano Paluri spent more than a decade building Meta’s AI efforts. Warp is building a US-based chemistry contract research organization that combines chemists, AI, and robotics. Biotech companies have outsourced much of this labor-intensive synthesis work to contractors in China such as WuXi AppTec, and Warp wants to bring it back to the US.

AI finds a bug in 70% of the web. Researchers at Calif, Wiz, OpenAI, and Quarkslab disclosed two PHP vulnerabilities, tracked as CVE-2026-109618 and CVE-2026-109619 and nicknamed “php2shell”. Both are critical remote code execution flaws rated 9.2 on CVSS, and Palo Alto Networks found more than 22.6 million exposed instances that may be vulnerable. The flaws had sat in PHP for years. Wiz’s writeup argues that as AI cuts the effort needed to build complex exploits, vulnerabilities that once looked impractical become viable targets. In the same week, Reuters reported on a hacker who broke into websites at more than 100 companies, leaving almost all of the work to AI agents.

Salesforce buys Braze for $5.66B. Salesforce is acquiring Braze, whose tools let brands send personalized messages and respond to customer behavior in real time. Braze shares jumped more than 50%. This is Salesforce’s third large AI-era deal, following $3.6B for the AI agent platform Fin earlier this year and $8B for Informatica last year, as investors worry that AI adoption will erode demand for traditional software. Like Workday’s purchase of C1.ai last week, it shows incumbents buying their way into the AI-native stack instead of building it.

HARD MARKETS

The AI credit scare reaches the hyperscalers. Meta beat revenue expectations in Q3, but investors focused on the cost. Capex was $34.9B for the quarter, free cash flow turned negative, and Meta plans a new class of non-voting shares to help fund its AI buildout. The stock fell 7% on Thursday, and the Nasdaq dropped 2.6%, its worst day since June. The bond market is reacting too. Global corporate credit spreads hit a 17-month high, and Meta’s credit default swaps reached a record as investors brace for a wave of debt sales to pay for the next round of data centers. Last week the stress was mostly on neoclouds. This week it reached the biggest buyers of compute. Citadel’s Ken Griffin told a conference that investors are “almost certainly” in an AI bubble.

Anthropic is the only big tech IPO left this fall. Switch, Nscale, and SB Energy have delayed their listings, and AirTrunk REIT and Firmus have pulled or paused theirs. That leaves Anthropic, which is pressing ahead with a mid-November IPO and plans to start marketing shares the week of Nov. 9. Its draft prospectus shows about $16.3B in first-half revenue, with roughly 20% coming from two customers. It also shows an $80.7B net loss, $72B of which is accounting charges, and at least $518B committed to compute and infrastructure over the next decade. Some investors are arguing for a $1.5T valuation instead of the $2T Anthropic reportedly wants. OpenAI, meanwhile, is raising up to $30B at a $1.25T pre-money valuation, below the $1.4T Bloomberg reported a month earlier. As with Lambda two weeks ago, a large share of the revenue depends on a small number of customers.

The model price war. OpenAI’s Decisions API is now GA after a public beta that opened Oct 6. It runs on gpt-6-luna, costs $0.10 per 1M input tokens, and charges nothing for output or cache. It returns structured answers (a probability, a choice from fixed options, or a score) instead of text, which puts it in direct competition with TypeSafe’s Jev, which we covered last month. Pricing is already moving customers between labs. Replit has made OpenAI the default for Replit Agent, citing performance, economics and its partnership. That could sharply reduce its spending with Anthropic, because most customers stay on the default, and Replit says it was recently a top-five Anthropic customer. The labs are now competing on unit economics as much as on capability.

HOW TO STARTUP

Human attention is the new bottleneck. Anthropic published a report on how its own workforce has changed. By August 2026, the AI agents employees direct were working six times as many hours as the employees themselves. Agents now move so fast that 80% of the elapsed time in Claude sessions was spent waiting for a person’s input. Junior hires’ share of technical hiring fell from 27% to 9%, and hiring managers now weigh autonomy and judgment more heavily than specific technical skills. For founders, the lesson cuts two ways. Organizations need to be designed around supervising agents, not just using them, and tools that manage human attention look like a real product opportunity. “This acceleration makes human attention the bottleneck for agent output.”

The enterprise opening may be outside engineering. Uber has rolled out ChatGPT Work, OpenAI’s enterprise AI agent, to all of its roughly 30,000 employees after piloting it with 1,000. About 80% of the early users were not engineers. They used it to analyze operations data, localize training materials, and build dashboards. Uber says it will track how employees use the tool to guide AI deployment across the business. For startups selling agents into large companies, the demand is coming from operations, training and analytics, not only from engineering teams.

HOW TO VENTURE

The big get bigger. The Fund CFO reports that US venture funds raised $108.5B through Q3, and funds over $500M took more than 78% of it. Murph Capital’s Q3 report goes further: across 152 closes, emerging funds (Fund I or II under $250M) got 9% of the dollars despite making up 57% of the closes, while funds of $500M or more got 76% while making up 17%. Bessemer’s $5.75B close alone raised more than all 86 emerging funds combined. This puts numbers behind last issue’s “short kings” piece, as capital keeps concentrating at the top.

Cheap money is the moat. Wing’s Chris Zeoli argues that two companies running the same GPUs can pay 7% or 15% to finance them. On $1B of equipment, that gap is $80M a year, enough to decide who can cut prices and who has to raise them. Big tech borrows at about 0.6 percentage points over Treasuries, while neoclouds pay 8 to 12 points over, close to the riskiest junk bonds. His advice for infrastructure investors is to judge the financing plan as carefully as the technology: who the anchor customer is, how strong its credit is, and what the company would pay to borrow if that customer left.

Aquant
Senior Enterprise Account Executive (NYC)

Hornet
Senior Research Engineer (Trondheim, Norway)

Motorica
Product Marketing Manager (Stockholm)