Hyperion’s bullish case for the long-term economics of AI

Hyperion Asset Management, however, has no such qualms. In a new whitepaper, Evidence Regarding the Likely Long-Term Economics of AI, Chief Investment Officer Mark Arnold and Deputy CIO Jason Orthman argue that the returns will be materially higher than the cost of capital.

Many of the assumptions on AI capex and the revenue generation needed to turn a profit are anchored to current applications. Again, this isn’t unreasonable, after all it’s tough to model revenue projections for entire new industries that could emerge with any reliability. On the other side of the coin, however, it is also extremely unlikely that the uses of AI will be limited to what is available today.

“Each increment in AI capability – from chatbots that draft, to agents that act, to humanoid robots and autonomous vehicles that work in the physical world – raises the productivity of existing labour and decision-making, lowers costs to consumers and enterprises for existing products, and creates new products and value that did not previously exist,” Hyperion says.

“The applications that create the highest value sit at the frontier, not on last year’s model.”

The hardware market for humanoids alone, Hyperion notes, could run to many trillions of dollars in total addressable market if robots take even a partial share of physical labour.

The economic logic that follows is straightforward. As intelligence gets cheaper, more use cases become viable. Jevons' paradox applies: the inference cost of a GPT-3.5-level system fell more than 280-fold between late 2022 and late 2024, yet industry spend on AI compute rose, not fell. Hyperion points to Epoch AI estimates that the price of a given performance level has been falling at a median of around 50-fold per year since 2024.

The revenue projections

The scale of the opportunity Hyperion is describing is not modest. The firm estimates that frontier AI model providers could capture more than US$8 trillion of revenue within a decade. That’s a staggering number, but there is the potential that it may prove conservative: SpaceX's prospectus put the AI total addressable market at US$28.5 trillion, of which US$22.7 trillion is enterprise applications.

“As customers pay for these new useful and valuable products and services, the economic pie grows; it is not merely reallocated,” Hyperion says.

“In our view, AI will expand the economy in real terms, not merely reallocate it. This industrial revolution is only just beginning, and a lift in trend growth over the next decade should significantly enlarge key addressable markets – power infrastructure, generation and storage; compute design, fabrication and infrastructure; autonomous transport and manufacturing; and advertising and software, including AI.”

“We expect the market for intelligence – AI applications and, separately, the compute that produces it – to be the fastest-growing TAM in the new paradigm, and in the long term the largest, outpacing even a much stronger overall economy.”

Evidence of the trajectory is already appearing in unit economics. Anthropic's inference margins have expanded from the high-30s to around 70% in a single year, the company has moved into operating profit, with estimates placing its annualised revenue as over US$100 billion by end of 2026.

OpenAI's enterprise revenue has accelerated sharply, with its Codex coding agent reaching 15 million active users, up more than 15-fold this year. Frontier model providers, Hyperion argues, are not burning cash indefinitely on an uneconomic product, they are instead reinvesting into the next, more expensive training run.

“That spending is a temporary scale-up cost, not evidence that the mature product is uneconomic.”

Training versus inference: A shifting dynamic

One of the less-discussed dimensions of the AI build-out is how the economics of compute are changing as the mix shifts from training to inference. Training a new model is capital-intensive but episodic, while inference is recurring, grows with adoption, and is already profitable at the frontier.

Crucially, inference already accounts for around two-thirds of total AI compute and is growing faster than training.

“We believe that, as model capabilities continue to improve and compute capacity available to AI model providers increases, use cases grow exponentially, the level of compute that is associated with revenue-generating use cases (inference) increases, and the level associated with training decreases,” Hyperion says.

“This will lead to improvements in the economics that AI model providers can generate.”

The data centre landlords

The supply-side proof of Hyperion's thesis comes from the economics of the compute providers themselves, as seen in the chart below of payback periods for the “landlords who provide the compute to model providers”.

Amazon (NYSE: AMZN) is now a US$169 billion annual recurring revenue business with a backlog of US$496 billion growing at triple-digit rates. Its AI and custom-silicon business already exceeds a US$25 billion run-rate. Most importantly, Amazon expects to break even on a given investment in a little under three years, against a five-to-six-year useful life for chips and a 30-year-plus life for data centres, and most capacity is already contracted for at least five years. The capital is recovered inside the first customer contract, with subsequent server generations generating returns on infrastructure that has already been paid for.

Microsoft (NYSE: MSFT) surpassed US$100 billion of revenue for FY26, growing 41%, with growth ahead of expectations. Azure revenue growth is expected to accelerate to 45% in the upcoming quarter. Microsoft's distinctive advantage is the enterprise layer above raw compute: more than 11,000 models in its catalogue, 100,000 customers on its Foundry platform, and nearly 90% of the Fortune 500 grounding their agents in enterprise context through it.

Alphabet (NYSE: GOOG) grew revenue 82% year-on-year, with backlog jumping US$50 billion sequentially to US$514 billion. Customers are exceeding their existing commitments by more than 50%. AWS and Google Cloud both printed incremental operating margins – 52% and 47% respectively – well above their reported margins of 39% and 36%, demonstrating that AI-driven revenue is accretive to margin profiles.

SpaceX (NASDAQ: SPCX) delivered its first public accounts since listing. Its AI division – formerly xAI – grew revenue 247% year-on-year to US$2.56 billion, driven by cloud-services agreements with Anthropic and Google. SpaceX CFO Bret Johnsen said new AI capital deployments are earning payback in less than a year – shorter still than Amazon's sub-three-year figure, and well inside the five-to-six-year useful life of the GPUs.

CoreWeave (NASDAQ: CRWV) reported June-quarter revenue of US$2.6 billion, up 112% year-on-year, with adjusted EBITDA margins of 59% and a contracted backlog of US$104 billion. Its average cash payback is approximately 2.5 years. Independent neoclouds like Nebius are printing paybacks of less than two years on second-quarter deals, providing an external check that the economics work outside the hyperscaler bundle.

“The most striking common thread across the quarter is that every scaled provider is capacity constrained,” Hyperion says.

“Amazon, Microsoft, Alphabet and Meta each told the market that demand continues to exceed the compute they can bring online, and each raised capital expenditure guidance accordingly. This is not the picture of an industry struggling to find a use for its investment; it is one racing to keep pace with demand it cannot yet satisfy.”

The Hyperion thesis

“Backing those companies that innovate rapidly and are highly vertically integrated across the AI stack improves the probability of investment success.”

According to Arnold and Orthman, Hyperion has aimed to remain “process-consistent” and position their portfolios towards companies that will be the “largest beneficiaries of the AI structural shift over the coming decade”.

• Chip, manufacturing and intellectual-property businesses that underpin AI: Nvidia, Arm, ASML and TSMC
• Large-scale AI compute infrastructure owners: Amazon, Alphabet, Microsoft, Meta and SpaceX
• Emerging fields of autonomy, humanoid robots and space, including satellite-based connectivity: Tesla and SpaceX.

“Several of these names are vertically integrated across the AI stack, designing silicon, owning compute infrastructure and distributing intelligence through their own products. That integration is, in our view, a source of business robustness: it lowers the cost of intelligence, enables faster execution and innovation, and secures supply.

“It also allows more of the value created to be captured inside the franchise rather than leaking to external parties and reduces risk compared to only having exposure to a single layer of the stack.”

Ultimately, Hyperion wants to back the companies that will disrupt industries and actually benefit from AI.

“Hyperion assesses value over a 10-year horizon. We have raised our long-term valuations for the hyperscalers and other AI beneficiaries over the past year on the strength of the increasing evidence of the attractive and sustainable economics of the AI ecosystem,” Hyperion adds.

“The forecast internal rates of return of these AI related businesses are currently attractive on a long-term basis.”

You can read Hyperion's full whitepaper here.

ETF

Hyperion Global Growth Companies Fund - Active ETF (HYGG)

Global Shares

View