SpaceX’s IPO has a way of pulling the spotlight toward the most visible kind of risk: rockets, launch cadence, manufacturing bottlenecks, regulatory approvals, and the relentless physics of getting payloads to orbit. But there’s another risk running alongside it—less cinematic, harder to measure, and increasingly central to how investors think about the future. It’s the risk of not knowing the economics of frontier AI.
The connection may sound forced at first. SpaceX is a company with metal, fuel, and flight hardware. Frontier AI companies are built around data pipelines, compute clusters, and model training runs that can cost millions before a single user pays. Yet the underlying question is strikingly similar: what does “scale” actually buy you, and when does it stop being a promise and start being a durable business?
In both cases, the market is trying to price something that is still being invented in real time. With SpaceX, the invention is industrial: reusable rockets, satellite manufacturing, and a vertically integrated supply chain. With frontier AI, the invention is economic: how value is created across training, inference, distribution, and ongoing updates—and whether the cost curves that look plausible on paper hold up once the world starts using the models at scale.
That’s why the quiet parallel story matters. Even as frontier AI models keep improving quickly, the basic business model remains unsettled. We still don’t have a universally accepted answer to what sustainable returns look like for the companies building the most capable systems, or for the platforms distributing them. The uncertainty isn’t just about demand. It’s about the structure of costs and the mechanics of monetization.
To understand why, it helps to break down the frontier AI stack into its economic components and ask what each one does to margins over time.
Training: the upfront bet that doesn’t behave like traditional R&D
Training is often described as “one-time” spending, but that framing hides the reality that frontier models are rarely truly one-and-done. Even if a company trains a flagship model once, it typically needs follow-on training cycles, fine-tuning, continual improvements, and safety or alignment work that evolves as capabilities and usage patterns change. In practice, training behaves more like a recurring investment program than a single project.
The cost drivers are also not stable. Compute prices can shift with energy costs, GPU availability, and supply chain constraints. Data quality and curation methods evolve. Model architectures change. And the industry’s understanding of what scaling laws imply for performance per unit of compute is still being refined. That means the “cost to reach a certain capability level” is not a fixed number—it’s a moving target.
Investors can tolerate uncertainty in early stages, but they want to know whether the uncertainty narrows as the company matures. For frontier AI, the narrowing is not guaranteed. Training costs might fall due to better efficiency, but they can also rise if the industry keeps pushing toward higher capability thresholds that require disproportionately more compute. The result is that training economics can look like a treadmill: every time you get closer to a stable margin profile, the bar moves.
Inference: where the money is supposed to be made, but where the bill can surprise you
If training is the upfront bet, inference is the ongoing toll. Every query, every token generated, every tool call—these are measurable costs. The challenge is that inference costs don’t map neatly onto revenue unless pricing and usage patterns are well understood.
Many frontier AI businesses monetize through subscriptions, usage-based pricing, enterprise contracts, or API calls. But the relationship between “how much a customer uses” and “how much it costs to serve” depends on several variables that are still not fully standardized across the industry:
1) Output length and complexity: A model that produces longer responses or performs multi-step reasoning can dramatically increase token consumption.
2) Latency requirements: Faster responses can require more expensive infrastructure or more aggressive optimization.
3) Tool use and retrieval: If the system calls external tools, searches databases, or retrieves documents, the cost profile becomes a blend of model inference plus orchestration overhead.
4) Caching and batching: Some workloads can be optimized heavily, while others cannot.
5) Model routing: Companies may use smaller models for simpler tasks and reserve frontier models for complex ones. This can improve margins, but it introduces new engineering and product complexity.
The industry has learned a lot about inference efficiency, including quantization, distillation, speculative decoding, and better batching strategies. But the key point is that these improvements don’t automatically translate into predictable margins. They can be offset by rising demand for more capable outputs, more frequent usage, and more demanding enterprise workflows.
In other words, inference economics can improve, but they can also be “taxed” by the very success of the product. When customers adopt the model for more tasks, they often ask for more thorough answers, more context, and more automation. That increases usage intensity, which can keep costs rising even if per-token costs fall.
Distribution: the hidden lever that determines who captures value
A frontier model is not a business by itself. It becomes a business when it is distributed—embedded into products, sold through channels, or offered via APIs. Distribution is where many companies discover that the economics are not only about technology; they’re about power.
If a model provider sells directly to end users, it can capture more of the value chain. But direct sales require go-to-market capabilities, support, compliance, and often custom integrations. If a model provider relies on partners—cloud platforms, app developers, enterprise software vendors—the partner may capture a larger share of the economics, leaving the model provider with lower margins.
This is where the analogy to rockets becomes useful. SpaceX’s advantage is not only that it can build rockets; it’s that it can manufacture, launch, and sell services in a way that reduces friction and improves reliability. In AI, the equivalent advantage is not only that a company can train a frontier model; it’s that it can distribute it in a way that makes adoption easy and value capture durable.
Yet distribution is still evolving. Many enterprises are experimenting with AI pilots, and procurement cycles can be slow. Meanwhile, consumer adoption is volatile: users try tools, churn, and return depending on perceived usefulness. The result is that distribution economics can swing widely, making it difficult to forecast long-term returns.
Ongoing updates: the recurring cost that doesn’t show up in the initial pitch
Frontier AI models face a problem that traditional software companies recognize immediately: the world changes, and users’ expectations change with it. Models need updates for performance, safety, and relevance. They also need monitoring to detect failures, drift, and misuse.
This creates a recurring cost base that is often underestimated in early discussions. It includes:
– Safety and evaluation pipelines
– Red-teaming and adversarial testing
– Data governance and compliance
– Customer support and incident response
– Product iteration based on user feedback
Some of these costs scale with usage; others scale with the number of deployments and integrations. Either way, they add to the “keep the lights on” burden. The question for investors is whether these costs become more efficient over time—through automation, better tooling, and standardized evaluation—or whether they remain stubbornly high because each new deployment environment introduces unique risks.
Pricing: the biggest unknown because it depends on behavior, not just capability
Pricing is where the frontier AI business model is most unsettled. Capability improvements do not automatically translate into willingness to pay. Customers pay for outcomes, not benchmarks. And outcomes depend on how the model fits into workflows.
A model that scores well on tests might still fail to deliver business value if it is unreliable, hard to integrate, or too expensive to run at the required frequency. Conversely, a model that is “less frontier” might win if it is cheaper, easier to deploy, and good enough for a specific task.
This is why pricing strategies vary widely across the industry: some charge per token, some per seat, some per workflow, and some bundle AI capabilities into broader software packages. Each strategy implies different assumptions about usage intensity and cost structure.
The open question is whether pricing will converge toward a stable equilibrium. In many markets, competition eventually forces pricing to align with marginal cost plus a reasonable margin. But AI is unusual because marginal cost is not the only factor. There is also the value of differentiation, brand trust, and ecosystem lock-in.
If a company can create a strong ecosystem—tools, integrations, developer platforms, and enterprise relationships—it may sustain pricing power longer than marginal cost would suggest. But if the market commoditizes model access, pricing could compress quickly. The industry is still deciding which path it is on.
Cost curves: the promise of scale versus the reality of constraints
When people talk about AI economics, they often focus on cost curves: the idea that as companies scale, they get cheaper per unit of output. That’s true in principle. But cost curves in frontier AI are shaped by constraints that don’t behave like simple economies of scale.
Compute is constrained by hardware supply, energy availability, and scheduling. Data is constrained by quality and licensing. Talent is constrained by expertise in training, optimization, and safety. And distribution is constrained by integration complexity and procurement realities.
So even if a company improves efficiency, it may still face bottlenecks that prevent costs from falling as fast as expected. Additionally, the industry’s competitive dynamics can push companies to spend more rather than less. If rivals are racing to match or surpass capabilities, the incentive is to invest, not to optimize margins.
This is where the “unknown” becomes more than a vague concern. It becomes a structural feature of the market. Frontier AI is not just a product category; it’s an arms race with economic consequences.
The investor’s dilemma: how to underwrite a business when the rules are still forming
An IPO is, in part, a bet on predictability. Public markets reward companies that can demonstrate repeatable economics. But frontier AI companies are still learning what repeatable economics look like.
That doesn’t mean the businesses aren’t real. It