How AI’s Boom Could Bankrupt Its Builders

OpenAI’s recent podcast hinted at a stark truth: cheap intelligence fuels demand, but also shreds the margins of the startups chasing it.

Welcome to the latest edition of Buy the Rumor; Sell the News. We’re closing in on 2,000 subscribers, including institutional allocators, venture capitalists, litigators, senior executives and entrepreneurs. Thank you to all who have subscribed!

In today’s post, I take a look at a recent podcast interview featuring two OpenAI execs, Brad Lightcap, its COO, and Ronnie Chatterji, its chief economist.

Separately, I provide my own commentary about an important point Brad raises, which is that inference usage scales super-linearly as inference price drops. This has devastating consequences for startups building on top of LLMs, and the venture capitalists who invest in them.

If you want to connect with me directly, my contact information is at the end of this post.

---

What OpenAI’s Top Brass Are Really Saying

OpenAI’s latest podcast episode offered a surprisingly transparent look at how the organization thinks about the deployment of artificial intelligence and where they believe its economic and social impact is headed. Their conversation exposed the inner scaffolding of OpenAI’s strategy and hinted at where the biggest disruptions might land first.

From playground to paradigm shift

Brad Lightcap traced how ChatGPT went from a developer’s playground experiment (the original text completion API) to a ubiquitous consumer phenomenon. The “big unlock,” he argued, wasn’t necessarily GPT-4’s raw intelligence but rather ChatGPT’s conversational interface, which resolved the blank-canvas problem and revealed pent-up demand.

Users who didn’t know what to do with a text-generation playground immediately knew how to interact with a chatbot. This re-framing of the user experience was, in Lightcap’s view, more pivotal than the subsequent improvements in model quality.

A two-pronged deployment mission

Lightcap described OpenAI’s mission as fundamentally twofold:

1. Build the most capable AI systems (the research engine), and

2. Deploy them safely and beneficially into the world (the operational mandate).

His own role is to figure out how to integrate these systems into real-world workflows, with all the complexity that entails, including cross-country norms, industry compliance frameworks, and psychological user expectations.

Chatterji’s role as chief economist is equally strategic: he’s mapping where, how, and at what speed AI-induced economic transformations will play out. This isn’t mere ivory-tower econometrics. OpenAI wants to identify which sectors and geographies will feel the impact first, and to translate that knowledge into actionable insights for governments and businesses worldwide.

The first-order use cases

So where does OpenAI see the highest near-term leverage?

• Software engineering: Tools like Cursor and Windsurf promise to 10x productivity, not merely by autocompleting code, but by eventually handling QA, writing unit tests, and orchestrating entire build pipelines. Given that the world’s biggest constraint on software output is human expertise, this is an obvious first beachhead.

• Scientific research: Think of drug discovery and materials science as endless corridors of locked doors. Traditional science forces researchers to choose only a few doors to open. AI tools can “peek behind them all,” vastly accelerating hypothesis generation and experimental iteration.

• Professional services: Private equity, investment banking, consulting: fields with high-margin, slide-heavy, memo-heavy deliverables are ripe for automation. OpenAI anticipates an entire retooling of these industries, where partners can focus on judgment and relationships while AI handles vast swaths of the analytical grunt work.

• Small businesses and emerging markets: In Africa, Chatterji noted, agricultural extension services (guiding farmers on seeds and fertilizers) have life-changing ROI but are too scarce. AI advisors could democratize this expertise. Likewise, a restaurateur or small manufacturer anywhere can now ask a GPT for tailored business advice, which was previously the domain of expensive consultants.

---

The deeper thesis: price elasticity of intelligence

Lightcap offered a simple but profound observation: when OpenAI cuts the cost of its models, consumption of “intelligence” skyrockets. There is apparently no near-term saturation point: heap tokens beget more tasks, which beget more demand. This mirrors a classic Jevons paradox: making a resource cheaper often increases total consumption dramatically, not less.

Chatterji framed this as potentially wonderful. Lowering the marginal cost of cognition might enable vast new categories of people and businesses to participate in legal, financial, and technical ecosystems. In other words, AI could become the ultimate democratizer of expertise.

Why this should worry startups and their investors

Here’s the uncomfortable corollary: the price elasticity of intelligence doesn’t just drive usage. It also threatens to destroy margins. That’s because Jevons-like demand rebounds often lead to booming aggregate volumes but shrinking per-unit profits. And that’s exactly the trap many LLM startups, and their venture backers, seem to be ignoring.

The Jevons paradox for cognition

When the steam engine improved coal efficiency, 19th-century economists expected total coal demand to drop. Instead, it soared, because cheaper power created new industries. Likewise, as token prices fall, people find more and more ways to embed AI into workflows. The total compute load on clusters grows faster than individual tasks get cheaper.

That’s fine if you’re the entity selling compute (like OpenAI, Microsoft, Google). But for application startups that build on these APIs, this dynamic is perilous. Their costs (per token) might decline gradually, but the market price for LLM calls falls even faster under competitive pressure.

The collapsing gross margin problem

A typical LLM app’s economics look like this:

• Revenue: $X per 1,000 tokens (often under fierce downward pricing pressure).

• Cost: ~$Y per 1,000 tokens (as billed by OpenAI or Anthropic).

But because demand is extremely elastic, meaning usage explodes as prices drop, these companies quickly find that:

1. They’re forced to slash prices to compete, destroying per-token revenue faster than their unit cost falls.

2. They can’t make it up on volume because there’s a hard ceiling on how much each user will pay for a marginal increase in output, even if they consume more.

In effect, they’re caught between hyperscalers racing to commoditize intelligence and a customer base trained to expect ever-cheaper thinking. (As an aside, it’s worth noting that as the model makers (OpenAI and its competitors) converge on the same set of frontier LLM capabilities, they’re all moving down the stack into infrastructure. Witness OpenAI’s $500 billion Stargate project, Anthropic’s compute cluster that it’s building with Amazon, etc. )

Meanwhile, a startup still has infrastructure costs: GPUs, vector databases, orchestrators, customer support. That means a startup’s net contribution margin can actually deteriorate even as aggregate demand booms.

Why do so few VCs or founders see it?

There are three systemic blind spots:

• Optimism bias from software history: Entrepreneurs and VCs tend to over-index on past SaaS narratives, where higher usage usually led to fatter margins. But unlike traditional SaaS, LLM apps often pay out a real-time variable COGS for every user action. They’re closer to digital utilities than to zero-marginal-cost software.

• Blind faith in “moats”: Many pitch decks promise data network effects, proprietary retrieval layers, or specialized fine-tunes. But until those genuinely lock out competitors (which they rarely do early on), everyone’s still paying roughly the same to OpenAI, so margin pressure is systemic.

• Failure to model hyper-elastic demand: Most founders still build decks showing straight-line growth in ARPU. In reality, the more elastic your demand, the more it explodes in usage while simultaneously cratering in revenue per unit, especially if competitors are willing to race to the bottom.

How to avoid the trap

The best founders are steering away from per-token business models toward:

• Outcome-based pricing: Charge for results, such as an approved insurance claim, a bug squashed, or a hire placed, where AI is merely a cost component, not the billing unit.

• Deep vertical integration: Build proprietary data or industry-specific workflows that make the GPT calls just a tiny slice of delivered value.

• Hybrid infrastructure: Offload routine calls to local quantized models; reserve expensive API hits for critical reasoning.

The uncomfortable conclusion

The elasticity of intelligence is a double-edged sword. It guarantees society will keep inventing new use cases. This is great for the economy. But it also means the next OpenAI price cut (or the next open weights breakthrough) could wipe out hundreds of fragile LLM startups that never found a moat beyond clever prompt engineering.

In that sense, OpenAI’s podcast was almost a polite warning. They’re building a world where intelligence is “too cheap to meter.” Entrepreneurs should plan accordingly by owning the customer relationship, the context, or the outcome, not merely by relabeling someone else’s cognitive commodity.

---

Coda

If you enjoy this newsletter, consider sharing it with a colleague.

Most posts are public. Some are paywalled.

I’m always happy to receive comments, questions, and pushback. If you want to connect with me directly, you can:

• follow me on Twitter,

• connect with me on LinkedIn, or

• send an email to dave [at] davefriedman dot co. (Not .com!)