The State of AI: Models, Moats, and the Consumer Renaissance

Summary

• Anish Acharya plants his flag in the “many winners” camp on frontier labs. In the last 2 weeks, xAI went “from not even being a real contender on the model side to being one of three” — a two-horse race became three — while Anthropic’s apparent dominance gave way to OpenAI’s excellent 3 months (new models, the Codex harness, and the ChatGPT desktop app), with multiple labs growing despite each other’s successes. Jen adds the tradeable overlay: X is an imperfect weather vane but an early indicator of developer sentiment; Claude is taking token-usage pushback, and Anthropic is going public later this year.
• The under-discussed macro scenario isn’t the bubble — it’s “what if we’re insufficiently optimistic?” B200 per-hour prices are rising even though it is a non-cutting-edge GPU and compute is normally deflationary, which “points to very constrained supply and essentially infinite demand.” On SaaS: February’s 30–40% drawdown was overselling and many names are back up 40%, but with SBC distortions now visible it’s “accelerate or die.”
• Most moats survive abundant low-cost intelligence — but the integration moat is at risk. Network, scale/distribution, and brand effects are “as good as they’ve ever been” (“no amount of coding agents is going to make Nike not Nike”), while coding agents make SAP-style integration dramatically better and raise an “existential question” for SIs and GSIs.
• Token spend rationally splits by bounded versus unbounded upside. For sales and product, “it’s economically rational to pay almost any price for a model that’s even 1 IQ point smarter” — “your Fable 5 or your Gro or your GPT56”; for finance, “you can’t close the books 10 times better than accurately,” so open-weight models plus reinforcement learning may be the Pareto-efficient choice. Models aren’t commodities: neurotic, literal GLM 5.2/5.3 versus open, presumptuous Kimi K3 — organizations need both minds.
• Labs are vertically integrating down into inference, not up into apps — inverting the early-2025 panic. Anthropic’s legal “plugin” (just collections of long prompt files) sparked a panic in which Thomson Reuters and other legal names traded down, but inference workloads are homogeneous and scalable while the app layer is idiosyncratic and OpEx-heavy; and in a multi-model Pareto-frontier world, labs have a harder time capturing “100% of your gross margin.”
• The consumer’s quarter may finally be here, but “we’re in the DOS era of AI” awaiting its Windows. Open-weight models are making AI software dramatically cheaper and more performant (Jen’s own X-timeline app cost $250 to onboard a new user), there’s still no AI-native app store, and consumers are excited to try and pay for new software — Anish compares the moment to Christmas 2009 with the iPhone, except consumers may now pay $200 a month.
• The investing posture has hardened around live product and technical founders. It’s now “disqualifying to not be showing a live product in a pitch at any stage”; founders skew “less MBAs, more researchers”; and per Ben at the offsite, “the biggest risk in the past was the ideas were too big and now the biggest risk is that the ideas are too small.” New business formation is at an all-time high outside a peak moment during COVID — the 25-year-old who would have been a YouTube creator now builds neighborhood SaaS.

Deep dive

1. Grok Bots buys the jeans — and the macro says demand is effectively infinite

• The opening anecdote that frames the episode: Anish mostly wears Frame jeans, photographed his current pair, told Grok Bots overnight “don’t spend more than $500 and get it done,” and woke up to a researched, purchased, in-transit pair — same fit, different wash, bought with his credit card. His read: “the defining characteristic of Grok Bots is resourcefulness,” and the next unlock comes from resourcefulness plus product architecture consumers understand.
• On the winner question, he’s “many winners”: xAI went from non-contender to one of three in 2 weeks; Anthropic went from seeming dominance to OpenAI’s excellent stretch — exceptional new models, the Codex harness, and a “very well done” ChatGPT desktop app — with specialization diverging and multiple labs growing. Jen’s caveat: X is an imperfect “weather vane” but an early indicator of developer sentiment, Claude faces token-usage pushback, and Anthropic goes public later this year.
• The bubble case is “fully discussed”; the out-of-distribution topic is insufficient optimism. Evidence: B200 per-hour prices rising even though the non-cutting-edge GPU is normally part of a highly deflationary market — “essentially infinite demand and highly constrained supply.”
• The SaaS whipsaw as market psychology: after February’s 30–40% drawdown, a16z said software was oversold; many names recovered 40% (“I’m not quite sure what we collectively accomplished”). Software is only 8–12% of enterprise spend, so the upside of vibe-coding your own payroll or CRM is low and the downside “essentially unlimited” — but the tide has receded on SBC-distorted economics, and those companies must “accelerate or die.”

2. Moats mostly survive; token spend splits by bounded vs. unbounded upside

• Working from 7 Powers, Anish argues most moats are untouched by abundant intelligence: “no amount of coding agents is going to make Nike not Nike,” and Instagram’s power was never the app’s complexity. For him, the most obvious exposed moat is integration — SAP is so complex that even migrating from one version to the next can be existentially risky, and SIs/GSIs face an existential question about their value as the integration point.
• The rational enterprise architecture: alpha-creating functions (product, sales, engineering, research) get frontier tokens because upside is unbounded — pay almost any price for 1 more IQ point, “your Fable 5 or your Gro or your GPT56.” Supporting functions have bounded upside: “the best way to close the books is accurately. You can’t close it 10 times better than accurately.” For those use cases, open-weight models with reinforcement learning can make sense on the Pareto-efficient cost curve.
• Jen surfaces Decagon founder Jesse Zhang’s post that for many startups open source is “actually the only option” — for localization, training, and fine-tuning, not just cost. Anish’s addition: reinforcement learning on a specialized problem and its reasoning traces builds compounding domain advantage, with Harvey seeing strong results in legal. The trade-off is lost generality — fine for legal or customer support.

3. Models are not commodities — and labs integrate down, not up

• His Big Five framing: there are “autistic models” — GLM 5.2 and 5.3 are “very literal and they’ll only do exactly what you told them” — versus Kimi K3, “open and presumptuous and creative.” You can’t be both highly open and highly neurotic; accounting may want neuroticism, while design may want openness, so organizations need multiple minds.
• The legal-plugin panic as case study: Anthropic’s plugin — “a ZIP of skill files… just long prompts” — prompted a panic in which Thomson Reuters and other legal names traded down dramatically. But instead of integrating up, labs have integrated down into inference and compute, where homogeneous workloads allow enormous scale; the app layer’s idiosyncratic pricing, packaging, and buying behavior make it “OpEx-heavy” for labs.
• Specialization is already visible in harnesses: OpenAI’s new GPT models and desktop app are a strong product container for knowledge work; Claude Code, terminal UI and all, is oriented to software engineering. Aggregation can beat any single lab — the Expedia metaphor: Cursor uses a frontier model for planning and a lesser one for execution; creative tools combine ElevenLabs (voice and music) with Black Forest Labs (video and creative direction); research can run the same query adversarially through models with non-overlapping training data and then use another model to converge. Labs are structurally limited to providing their own in-house models, while an application aggregator can provide best of breed.

4. Apps productize the intelligence primitive; loops are the enterprise unlock

• The core analogy: intelligence is a primitive like cloud, and just as Salesforce turned the AWS cloud primitive into CRM, “you really need Harvey to turn that into an economic outcome for the legal industry.” Credit unions illustrate the idiosyncrasy — most don’t want to halve headcount; they want to double it while remaining economically performant.
• Agent demystified: “just a model in a loop with tools and memory.” The coding loop — bug reported, reproduced, fix generated, verified, shipped autonomously if low-risk — extends to price optimization and procurement, up to the business loop where the model proposes, “I think we need to open a branch in Tijuana.”
• Marc’s line is “industries, not markets”: the coding primitive spans Claude Code exposing “the raw horsepower” to developers, up to Replit abstracting for the non-coding small-business owner — all variations of pricing, productization, and packaging. On whether labs will squeeze the app layer: in 2023’s one-model world “the labs would just take 100% of your gross margin over time”; with many options across the Pareto frontier, they have a harder time doing that.

5. The consumer renaissance: DOS era, personal agents, and Town’s compounding memory

• What held consumer back: consumers don’t love paying for software while AI carries real marginal costs of distribution and engagement — Jen’s X-timeline app cost $250 to onboard a user, making a mass-market free product hard to support; open-weight models are now changing that with lower costs and better performance. There’s no AI-native app store, so distribution looks like Web 2.0, not mobile; and “we’re in the DOS era of AI… we’re going to need the Windows.”
• What’s working: coding agents for the “digitally native entrepreneur” — the YouTube-creator moral panic reframed as kids wanting to build internet businesses, now able to build $100K–$1M/year “mom-and-pop SaaS,” not venture-backable but “very cool for the country.” Personal agents are moving from OpenClaw’s January “Homebrew Computer Club” energy into consumer software such as Grok Bots and ChatGPT Work.
• His consumer definition: if you can’t justify sales-led acquisition, usually around a $15K ACV, it’s consumer — the plumber counts. Entertainment will be massive (Character was arguably an entertainment company; Asian short-form drama is starting to come over): “most people want to spend time, not save time.”
• The broader personal-agent model is a set of life loops — family, friendships, money, and health — in which changing information leads to decisions, agency, execution, and another iteration. OpenAI is focused on health and finance, while startups are working on shopping; Anish expects a dramatic quality-of-life improvement, with 80% of the surplus delivered to the mass market.
• Town (Alex Rampell’s investment) as the memory-compounding pattern: day 1 it’s a new hire; by day 30 it makes excellent assumptions from soaked-in context — showing up as retention and per-customer pricing power. Jen says her professional inbox is at zero but her personal inbox has about 20,000 messages; Town surfaces what matters, scrubs subscriptions, and is starting to self-improve by showing the credit cost of routines and how to save credits.

6. Economics, founders, and where the capital goes

• The consumer whitespace includes AI in “the emotional, interpersonal domain”: you can talk to Claude, OpenAI, or K3 “and feel feelings,” after “40 years of technology that boosted our intellect… but nothing that spoke to our humanity.” Startups can pursue products that labs and big tech are not culturally set up to build, such as a companion that disagrees with you or uses sexual innuendo. Consumers are excited to download new software and, unlike the 99-cent era, Anish says they are willing to pay $200 a month.
• Funding posture: mostly no pre-revenue bets — “it’s disqualifying to not be showing a live product in a pitch at any stage these days because it’s so trivial to build stuff”; the work is extrapolating from statistically significant sales and product against price and implicit risk, with small “pre-everything” call options for talented, experienced teams. The old wisdom that too much seed capital wrecks companies is becoming more nuanced: a focused $100M can deliver a different value proposition than $20M — a better problem than fintechs “indirectly subsidizing their customers through weak underwriting.”
• Founder archetype shift: “less MBAs, more researchers” — lower business sophistication, dramatically higher technical sophistication, which is “upstream of all the good things”; business sophistication can be taught and observed, while technical sophistication typically cannot. Ben’s offsite line: the biggest risk used to be ideas too big; now it’s ideas too small.
• For existing SMBs, the old channels remain, but founders increasingly need an original network effect — word of mouth — because Instagram, TikTok, and X make it difficult to build a new distribution channel on top of an existing one. The most interesting segment is new business formation, at an all-time high and highest outside a peak moment during COVID: “not the 55-year-old plumber… a 25-year-old building SaaS for their neighborhood or their high school.”