AI Has An Infrastructure Problem · Token ROI · Culture As Filter

AI is moving from model race to industrial bottleneck

Andrew Feldman’s case for Cerebras was not really a chip pitch. It was a claim about where AI competition has moved. He argued that AI infrastructure is being built behind demand, not ahead of it, with Cerebras sitting on a claimed $25bn backlog while data centres, power, fabs and memory constrain the whole system. NVIDIA’s $81.6bn quarter, roughly $56bn of profit, $91bn guidance and $80bn buyback made the same point from the other end of the stack: AI capex is no longer a speculative side bet. It is becoming a recurring industrial bill.

The contrarian read is that the bubble debate may be asking the wrong question. If demand is real but the physical stack cannot expand quickly enough, the problem is not overbuild; it is scarcity mispriced as hype. The next shortage may not be GPUs in the narrow sense. Feldman kept pointing to high-bandwidth memory, power and grid access. HBM supply is concentrated in Samsung, Micron and Hynix, fabs take years and tens of billions to build, and Micron’s reported 80–85% gross margins on memory look less like a cyclical gift than a signal of bottleneck power.

For operators, this changes how AI should be budgeted. Compute, power and memory are not generic vendor costs if they determine what workflows you can ship, how quickly customers get answers, and whether your margins hold at scale. The useful exercise is not “what AI tools should we buy?” but “which bottleneck are we buying around: latency, throughput, capacity, or access?” Treat AI capacity the way a finance team treats revenue: best case, base case, constrained case. The companies that plan for scarcity will make better product and pricing decisions than the ones assuming costs simply fall on schedule.

Token economics are becoming a CFO problem

The sharpest shift in the OpenAI, SpaceX and NVIDIA discussion was from capability to payback. Rory’s framing was blunt: the next $2tn of AI infrastructure spend has to earn its keep. Uber and Microsoft saying developer productivity gains are hard to measure matters because these are exactly the companies expected to absorb AI most naturally. If they cannot cleanly attribute productivity lift, weaker-margin businesses will be even less forgiving.

That does not mean AI spend stops. It means the buyer changes. The experimental budget gives way to the CFO’s spreadsheet: cost per workflow, volume uplift, labour displacement, quality improvement, cycle-time compression, and pricing power. “AI ROI is not measurable yet” sounds like a sceptical line, but the more uncomfortable read is that it is a warning shot. Spend will continue where the marginal token creates obvious value; it will get capped where teams cannot explain what changed.

This is where the agent-infrastructure names become interesting. Exa was framed as structured web search for agents; OpenRouter as the model-routing layer that lets teams choose the right model for the task and cost profile. These are not glamorous in the model-lab sense, but they sit directly in the ROI path. If agentic workflows become real, the enduring products may be the ones that route, search, observe and meter usage well enough for a business to trust the spend.

The tactical consequence is simple: audit AI like a unit-economics problem, not like a software procurement line. For every workflow, know the token cost, human time saved, output quality, failure rate and customer value. Premium AI positioning only works where the payback is legible. If customers say they can tolerate doubled AI prices, ask whether that means the product is mission-critical — or whether they are barely using it today.

Speed is becoming the product, not a feature

Feldman’s most useful product argument was about latency. He claimed Cerebras is 15x faster for architectural reasons and cited a public benchmark where it ran Kimi K2 6.7x faster than the next GPU cloud. The specific number matters less than the broader claim: on hard inference and search problems, a slow answer is not a worse version of the same product. It is a different product.

This is easy to underweight because software teams treat speed as polish after correctness. In AI, speed changes behaviour. A three-minute answer can support an interactive workflow; a twenty-minute answer becomes a batch job, a curiosity, or something users abandon. Faster systems create more sessions, more retries, more exploration and more willingness to pay because they sit inside the user’s loop rather than outside it.

The contrarian read is that “best model” may become a weaker category than “fastest useful system”. Benchmarks still matter, but buyers care about the lived workflow: can the system respond while the user still has intent, can it run many attempts cheaply enough, and can it be trusted in production cadence? A slightly less intelligent system that is fast, reliable and integrated may beat a frontier system that makes the operator wait.

Product teams should therefore define speed in business terms. What does lower latency unlock: more conversions, more internal usage, shorter sales cycles, higher support deflection, greater operational trust? If there is no answer, speed is a vanity metric. If there is an answer, latency belongs in the roadmap next to features, pricing and distribution.

Enterprise AI is blocked by permission, not intelligence

A useful thread across the AI conversations was that enterprise adoption is less constrained by model quality than by institutional permission. Feldman said lawyers and CISOs are the biggest blockers, not messy data. The OpenRouter and Exa discussion pointed in the same direction: the enterprise value layer is not just better answers, but safer routing, repeatable workflows, controls and observability. Dan Loeb’s governance lens gives the same idea a boardroom version: institutions fail when incentives and accountability are unclear.

This is a better explanation for slow enterprise AI than the usual “data readiness” excuse. Many deployments are technically possible but institutionally illegible. Nobody knows who is accountable when the model is wrong, which data can leave the boundary, what the audit trail looks like, or how a regulator would view the workflow. The result is not a model problem. It is a permission problem.

The less obvious implication is that governance can become a product moat. Procurement-safe AI will not win every demo, but it may win the budget. Controls, policy enforcement, provenance, audit logs, model choice, fallback paths and narrow approved use cases sound boring until the buyer is a bank, insurer, healthcare company or regulated marketplace. In those markets, trust is not a wrapper around the product. It is part of the product surface.

For founders selling AI into enterprises, the play is to pre-solve the objection before the CISO or legal team turns up. Start with narrow workflows where the risk is bounded and the value is measurable. Make governance visible in the interface and in the sales process. The demo should not only show what the AI can do; it should show why the organisation can safely let it do it.

Distribution is back because AI compresses product advantage

Nico Laqua’s most useful point in the Corgi interview was not the office mattress or the seven-day work week. It was his claim that AI makes sales and marketing more important, not less. Product quality alone is no longer enough when features can be copied faster, internal tools can be vibe-coded more cheaply, and customers are drowning in similar-sounding AI promises. His line that B2B sales remains strong while B2B marketing is awful is crude, but directionally important.

The contrarian read is that the current AI wave may reward the best distribution machines more than the best demos. Model access, design patterns and product surfaces diffuse quickly. Trust, narrative, category position, buyer relationships and proof of ROI diffuse more slowly. That means GTM is not a tax on product-led purity; it is where the defensibility may live.

Corgi’s own story pushes the point further. The company has raised at a reported $2.5bn–$2.6bn valuation, operates in regulated insurance, spends heavily on Anthropic, and still frames speed of fundraising and sales execution as a core company signal. “Good companies get deals done” is partly bravado, but the operator version is useful: slow deals often reveal unclear value, weak urgency or poor stakeholder choreography.

For product and GTM teams, the lesson is to make distribution more quantitative and more opinionated. Explain ROI in plain numbers. Build category language customers can repeat internally. Invest in sales motions that create trust before procurement friction arrives. Use marketing where it compounds — credibility, education, authority, and market definition — rather than as a content treadmill. In an AI market full of plausible products, the company that helps the buyer believe and act may beat the company with the better feature list.

Culture is a selection mechanism, not a vibe

The Corgi interview was easy to dismiss as founder theatre: seven-day weeks, the founder sleeping in the office, a mattress on site, a 24/7 cafe opened for under $100k, and a claim that many of the first 30 employees had Corgi tattoos. But the useful read is not whether other companies should copy the extremity. It is that Nico uses culture as an explicit hiring filter. The operating system repels people as deliberately as it attracts them.

That is uncomfortable because most companies prefer culture to sound broadly appealing. Corgi is doing the opposite. It is saying the quiet part out loud: the business is optimising for intensity, speed and asymmetric upside, not for being a good default employer for many talented people. The visible symbols — cafe, office, rituals, pace — are not decoration if they make the company’s true demands harder to misunderstand.

The contrarian lesson is that “culture theatre” can be real leverage when it matches the actual operating cadence. The danger is when symbols advertise a tempo the company does not really sustain, or when extremity becomes a substitute for judgement. But if the business genuinely requires unusual speed, ambiguity tolerance and personal commitment, hiding that in the hiring process is expensive. Mis-sold culture creates churn, resentment and management drag.

Operators should decide whether they are optimising for breadth or intensity, then design hiring around that choice. Work trials, weekend exposure, real-time problem solving and transparent discussion of pace can be useful if intensity is the strategy. If the goal is a broader talent pool, build different rituals and guardrails. The mistake is pretending every company can have the same employee promise while demanding a different level of sacrifice.

Durable systems need depth, not just brilliance

Darren Farber’s “magazine depth” frame was ostensibly about defence: Tomahawks, HIMARS, ordnance stockpiles, replenishment capacity and multi-year procurement matter as much as any individual platform. His point was that flexible power only works if the system can sustain use under pressure. Dan Loeb’s investing stories made a parallel point in companies: high-status brands such as Sony and Sotheby’s can hide weak governance, stale structures and under-optimised assets for years.

The shared insight is that first-order strength is often overvalued. A military platform, a famous brand, a talented founding team or a strong product launch can all look powerful while the underlying system is thin. What matters under stress is replenishment: capital, talent, inventory, trust, governance, decision rights and the ability to keep improving after the first move.

The provocative read is that prestige often masks brittleness. Farber applied that to China’s scale and internal illegitimacy; Loeb applied it to boards and management teams that become too loyal to status rather than shareholders. In business, the same pattern shows up when companies confuse brand heat with resilience, or early traction with a repeatable operating machine.

The practical exercise is to audit for depth. Where would the business break if demand doubled, a key supplier failed, a senior leader left, a regulator intervened, or a competitor copied the product? Do you have enough runway, inventory, management capacity, data integrity, customer trust and governance clarity to sustain tempo? Strategy is not only about choosing the right hill. It is about having enough depth to keep fighting after the first push.