Every Other Return Is Already Priced In

If the good stuff is already in the price, where does the return come from?

A company reports a blowout quarter. Revenue up 18%. Margins expanding. The CEO says all the right things on the earnings call. The stock opens flat.

The analyst note that morning explains it in one line. “Results in line with expectations.” The company just delivered its best quarter in three years, and the market shrugged, because it had already assumed this would happen. The revenue, the margins, the optimism. All of it was baked into the price months ago.

This happens constantly, and most investors never stop to ask what it means. If the market already prices in what it expects, then delivering on expectations doesn’t generate your return. The return comes from somewhere else entirely.

The Expectations Treadmill

Stock prices don’t just reflect what a company earned last quarter. They reflect what the market believes a company will earn over the next several years, discounted back to today. That’s the core of discounted cash flow logic, and while the math gets complicated, the intuition is simple. A stock price is a bet on the future.

Different companies carry different expectations. Nobody expects a utility to grow earnings 15% a year. The market prices utilities for stability and a dividend. Tech companies carry the opposite characteristics. A stock trading at 40 times earnings isn’t priced for modest growth. It’s priced for something big. When the big thing happens, the stock doesn’t jump because it already moved months ago.

This creates what feels like a treadmill. Companies have to keep running just to justify the price they’re already trading at. When they fall short, the stock drops. When they meet expectations, the stock holds. The only thing that makes the stock meaningfully go up is delivering something the market didn’t see coming, and that something has a name.

The Residual Is Innovation

Innovation, in the broadest sense, is the unexpected advantage. A new product. A new business model. A new way of doing something that the market hasn’t priced because it couldn’t be anticipated.

Consider Spotify. When it launched, digital music distribution already existed. iTunes had proven the model. You could buy any song for 99 cents. The market had priced in digital music. What it hadn’t priced in was a Swedish startup offering unlimited access to the entire catalog for $10 a month. That business model innovation created billions in equity value, not because the technology was new, but because the approach was.

Or think about the iPhone. In 2006, smartphones existed. Blackberry was dominant. The market had priced in mobile communication. What it hadn’t priced in was a device that would restructure the entire relationship between consumers and the internet. The idea was what generated the return, not the fact that phones existed.

Innovation is the most durable source of return that can’t be priced in advance. Everything else (interest rate expectations, regulatory changes, demographic shifts, industry trends) gets absorbed into stock prices as the market becomes aware of it. By the time you read about it, it’s in the price. Innovation, by definition, is the part that hasn’t happened yet.

Sparkline Capital put this well in their research on innovation investing. “In an efficient market, prices accurately incorporate future growth expectations. Even if a disruptive company does ultimately reshape society, its investors will not realize excess returns if this outcome was already priced in.”^[1] The return is the gap between what the market expected and what actually happened.

And the share of equity value that depends on innovation keeps growing. Intangible capital (intellectual property, brands, network effects, software) now comprises roughly 42% of the U.S. capital stock and is growing faster than tangible capital.^[2] The four largest companies in the world by market value don’t need significant net tangible assets to generate their earnings. Warren Buffett pointed this out in 2018. “They are not like AT&T, GM, or Exxon Mobil, requiring lots of capital to produce earnings.”^[2:1] The asset that matters is the one the market can’t easily anticipate.

This doesn’t mean markets are perfectly efficient. They aren’t. Behavioral biases, structural frictions, and information asymmetries create pockets of mispricing all the time. But the directional truth holds. Most of the expected stuff is in the price. Most of the return comes from the unexpected.

The Test Case

If innovation is the source of all true returns, then AI is the stress test.

The hyperscalers are spending hundreds of billions on AI infrastructure this year alone. The assumption embedded in those spending levels is that AI will generate productivity gains large enough to pay for itself many times over. That is a massive bet on innovation. The market is pricing in future AI-driven returns.

Three years after ChatGPT launched, the evidence for those gains is thin. Developer jobs haven’t disappeared. The products people use every day haven’t undergone a visible step-change in quality. AI may be producing gains in harder-to-measure dimensions (faster time-to-market, reduced headcount growth), but if those gains can’t be measured, they can’t be priced in either. And that only reinforces the point.

But part of the reason the innovation hasn’t materialized may have nothing to do with AI’s capabilities. It may have to do with how AI labs make money.

The Token Problem

The dominant revenue model for frontier AI companies is token consumption. You pay based on how much compute your queries consume. This creates a straightforward incentive. The more tokens an agent uses, the more revenue the lab collects.

Now consider the current race to build agents that can run unsupervised for hours at a time. The pitch is autonomy. Give the agent a task, go to bed, wake up to the finished product. That’s compelling. But think about what sits underneath it.

If an agent could deliver your result in one hour but instead ran for eight, you’d never know. You went to bed. The answer was there in the morning. The lab, meanwhile, collected eight hours’ worth of token revenue instead of one.

This isn’t an accusation. It’s an observation about incentive structure. When the revenue model rewards consumption rather than efficiency, the system naturally drifts toward more consumption. Current pricing is also subsidized by venture capital, which means the true cost of inference is obscured. The bet is that models will get cheaper to train and run faster than the subsidies dry up, so the price drop will eventually be organic. We’ll learn much more when these companies go public. Both OpenAI and Anthropic have reportedly discussed IPOs as early as late 2026, though neither has filed. Public filings will force the kind of transparency that private fundraising rounds don’t.

Businesses, historically, have grown in the increments defined by their era’s dominant technology. Ship routes. Factories. Web properties. If the current increment is tokens, then the incentive to consume more of them is the fundamental business model.

What Would Change the Assessment

If innovation is the source of all true return, and AI is the biggest innovation bet in a generation, then the key question is whether the innovation is real but early, or simply hasn’t materialized.

A few signals would shift the picture.

If AI-native startups (companies founded after 2023 that are built entirely around AI workflows) begin consistently outperforming established competitors on product quality and speed, that would be strong evidence that AI is generating genuine productivity gains, even if incumbents can’t capture them. This hasn’t happened yet at scale.

If industry-wide software quality metrics (defect rates, deployment frequency, mean time to recovery) measurably improve over the next 18 months, that would suggest AI is producing innovation surplus in development workflows. The data so far is inconclusive.

The domain where AI’s advantage seems most genuine is medicine, not software. Gene therapy and drug discovery involve pattern-matching at a scale that exceeds human cognitive bandwidth. DNA structures are essentially statistical patterns, and identifying therapeutic targets across millions of sequences is exactly the kind of problem where AI’s architecture fits. That’s entirely different from asking an LLM to write a React component. An Australian AI researcher named Paul Conyngham recently used ChatGPT and AlphaFold to help design a personalized mRNA cancer vaccine for his dog, working alongside university scientists who manufactured and administered it. The tumor reportedly shrank by 75%.^[3] The innovation there wasn’t the LLM replacing the scientist. It was the LLM making it possible for a non-biologist to navigate a domain he couldn’t have entered alone.

Here is the falsifiable version of this argument. If aggregate measures of software quality and worker productivity do not show a measurable step-change by the end of 2027, the “AI productivity revolution” thesis will need serious revision. Not abandonment. Technologies routinely cycle through false starts before finding their moment. Neural networks existed in the 1990s but were ahead of their time. Electric vehicles were the best-selling cars of the early 1900s before being eclipsed by internal combustion, and they didn’t reemerge for a century.^[1:1] The internet took 10 to 15 years to produce the productivity gains economists were looking for. But revision would be warranted, because the spending levels assume the gains are arriving much sooner than those historical timelines suggest.

The Innovation Discount

If innovation is the most durable source of true return, and the market is currently discounting innovation-heavy companies because of near-term pain, then a patient investor is effectively buying innovation at a markdown.

This has happened before. On June 1, 1932, the S&P 500 had fallen 86.2% from its 1929 peak, and a single company (AT&T) made up 12.7% of the entire U.S. stock market.^[4] An investor who bought at that moment of maximum pain would have earned roughly 16.1% annualized over the next 25 years, turning $100,000 into nearly $4.2 million.^[4:1] The point is not that today’s drawdown is comparable to 1932. It’s that the market’s harshest discounts on innovation have historically been followed by the strongest long-term returns.

Tech companies carry long-duration cash flows, meaning their projected earnings are heavily weighted toward the future. When interest rates rise, those distant cash flows get discounted more aggressively, and the stock price drops. This is mechanical, not analytical. It doesn’t mean the innovation is worth less. It means the market is applying a higher discount rate to future earnings.

The last 10 to 15 years created bloat in tech. Zero interest rate policy and pandemic stimulus let companies hire aggressively, spend freely, and defer profitability without consequence. That era is over. The companies that survive the current tightening will be leaner, more capital-efficient, and more focused. The hyperscalers themselves are on a trajectory toward utility status. Just as railroads became the platform on which industrial businesses were built, cloud infrastructure is becoming the platform for the next generation of startups. The returns from being a utility are real but moderate. The outsized innovation returns will come from whoever builds something on top of that infrastructure that nobody expected.

The private market pullback could help accelerate this. If AI genuinely reduces the time and cost of validating a startup idea (faster prototyping, cheaper user testing, simulated customer feedback), then smaller checks could fund leaner companies that reach proof-of-concept faster. The combination of cheaper validation and tighter capital could produce startups that are more capital-efficient than anything we’ve seen.

Or it could simply produce fewer startups. That’s the tension, and nobody knows which way it tips yet.

Where the Return Lives

Go back to the $100 stock. The company announced its best quarter in three years. The stock didn’t move. The investor who bought it expecting a beat earned nothing from the beat.

The return, when it eventually comes, won’t come from the expected growth that every analyst projected. It won’t come from the product launch that was on the roadmap. It will come from the thing that wasn’t on the roadmap. The business model nobody anticipated. The application nobody imagined. The efficiency gain that surprised even the company that achieved it.

That’s innovation. It’s the residual after expectations have been satisfied.

The market will make another all-time high. It always does. The S&P 500 has never failed to recover. The question for every investor isn’t whether to own stocks. It’s where the unpriced innovation is hitting, and whether you’re willing to hold it through the discount.

If this framework is right, the market is currently punishing many of the companies most likely to produce the next wave of unexpected advantage. That’s not a guarantee of returns, but a bet, and it depends on the assumption that innovation will continue to emerge from the same kind of capital-intensive, technology-driven companies that have produced it before. That assumption could be wrong. But the structural pattern is clear enough to take seriously.

This is general education and analysis, not personalized financial advice. The author holds no positions mentioned in this piece.

---

1. Kai Wu, Sparkline Capital, “Investing in Innovation,” April 2022.

2. Kai Wu, Sparkline Capital, “Investing in the Intangible Economy,” October 2020.

3. Fortune, “This startup founder used ChatGPT and AlphaFold to help treat his dog’s cancer,” March 15, 2026.

4. Jason Zweig, The Wall Street Journal, “The Big Scary Myth Stalking the Stock Market,” February 13, 2026.