ARK • Disrupt

Last week, earnings from Alphabet, Microsoft, and Amazon highlighted the continued surge in demand for AI and cloud infrastructure, prompting the hyperscalers to add to their investment in data centers, custom silicon, and computing capacity.¹ During the second quarter, cloud revenue growth accelerated at all three companies: Google Cloud from 63% on a year-over-year basis during the first quarter to 82%; Microsoft Azure from 39% to 43%; and Amazon Web Services (AWS) from 28% to 37%, its fastest growth in 18 quarters.²

Management commentaries also indicated that infrastructure remains a constraint as companies scale increasingly compute-intensive AI workloads. As a result, Alphabet raised its guidance for this year’s capital spending from $180–190 billion to $195–205 billion across its vertically integrated AI stack, from custom tensor processing units (TPUs), Axion central processing units (CPUs), and cloud infrastructure to Gemini’s frontier models. Amazon increased its capital spending guidance for 2026 from ~$200 billion to $220 billion, highlighting AI infrastructure investment and memory costs as well as semiconductors, robotics, and satellites.

Amazon gave the clearest indication of the imbalance between demand and available infrastructure. CEO Andy Jassy said that, even at ~$220 billion in spending, the company will not have enough capacity to meet demand, an imbalance that could persist into 2027. Amazon also reported that its AI and chips businesses had exceeded annual revenue run rates of $25 billion each and are growing at triple-digit rates on a year-over-year basis.

Vertical integration is becoming an increasingly important source of differentiation among the hyperscalers. Alphabet is combining proprietary TPUs and CPUs with Gemini and Google Cloud; Microsoft is integrating cloud infrastructure, both proprietary and third-party AI models, software distribution, and enterprise relationships; and Amazon is pairing AWS with Trainium, Inferentia, Graviton, Bedrock, and a broad selection of third-party models. Those integrated stacks should enable them to optimize performance, capacity, and cost across the entire AI computing system.

While questions abound about capacity utilization, depreciation, free cash flow, and the ultimate return on investment, the results thus far suggest that hyperscalers believe insufficient capacity—not excess capacity—is the immediate constraint.

After Starship Flight 13's soft splashdown, imagery showed cracked heat shield tiles and ablator streaks that suggested hot plasma penetrated tile seams.³ Last week, three veteran NASA thermal-protection engineers argued that Starship's ceramic tile heat shield is “a dead-end for all missions that require full and rapid reusability,” citing the painstaking inspection process necessary between each flight for the ~18,000 tiles.⁴ With history as their guide, they noted that inspections limited the Space Shuttle to four to five flights per year.

Its flight records suggest that SpaceX is making impressive strides toward heat shield reusability. Even as SpaceX tested Starship at higher dynamic pressure points to probe the shield limits, damage has trended down sharply, as shown below.⁵

Source: Musk 2026.⁶ For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security.

Unlike the Space Shuttle, which depended on human crews, SpaceX can afford to lose uncrewed Starships after launching satellites into orbit. Moreover, unlike the Shuttle on which one lost tile could have caused catastrophic failure, Starship's heat shield structure is much more forgiving.⁷ Each tile on the Space Shuttle's was unique, all of them taking two labor-years to attach and more than 17,000 labor-hours per flight for inspection, replacement, and maintenance. The project was born of blood, sweat, tears—and spit. Initially, the adhesive adhering to the tiles would dry so quickly that workers would spit on it to slow the process down, a practice NASA identified and discontinued.⁸ Starship's tiles are standardized and produced on an automated production line at a rate of one per 13 seconds, a fraction of the three days necessary to fabricate and install a single Shuttle tile.⁹ Recently, SpaceX completed a full tile strip-and-replace in two weeks, roughly four times faster than the Shuttle ever managed.¹⁰ To achieve rapid reuse, SpaceX is likely to continue streamlining the process.¹¹

Long before it reaches airline-like reusability, even single-digit reuses of Starship could have a meaningful impact on launch costs. If refurbishing tiles becomes the gating factor, compressing the process into days—equivalent to the fastest turn of the Falcon 9 booster—could lower the cost to orbit meaningfully from ~$1,000 today toward the $100-200/kg that Elon Musk has targeted.¹²

Over the past month, OpenAI halved the cost of frontier intelligence with the release of GPT-5.6, Anthropic released frontier-class model Opus 5 at cost per performance on par with OpenAI, and OpenAI lowered the price of Terra and Luna—its medium and smallest frontier-class models—by a further 20% and 80%, respectively. The cost of intelligence is dropping precipitously, as shown below.¹³

Source: ARK Investment Management LLC, 2026, based on data from Datacurve.ai as of July 30, 2026. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security or cryptocurrency. Past performance is not indicative of future results.

ARK’s Big Ideas 2026 illustrated that inference costs in 2025 declined by 99% year-over-year on a levelized performance basis. On agentic benchmarks, costs are now declining even more rapidly. At this year’s threshold of performance, 20% on DeepSWE, backward-looking costs dropped 7,000-fold—99.99%—starting from Anthropic's Sonnet 4.6 in February priced at ~$3 per task to OpenAI's GPT-5.6 Luna in July at $0.06 per task for the same level of performance. In the next year, on a levelized performance basis, costs should drop by another 99.96%, as shown below.¹⁴

Source: ARK Investment Management LLC, 2026, based on data from Datacurve.ai as of July 30, 2026. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security or cryptocurrency. Past performance is not indicative of future results.

Performance at a fixed level on the benchmark represents only one measure of the improvement in AI. On an agentic coding benchmark such as DeepSWE, 50% performance was not possible at any price point in early 2026. Now, that performance costs $0.15 per task. These increases in capability—at literally infinite cost declines—are unlocking markets: an AI model now can achieve what was impossible just six months ago. To illustrate this point, ARK derives a separate performance improvement curve on the deepSWE benchmark: the magnitude of error rate decline given an extreme number of chain-of-thought tokens. The top-end error rate on DeepSWE has been declining 4.2-fold at an annualized rate. Practically, at a cost of one dollar per task, the best-in-class model would have failed to deliver a successful result 92% of the time in March, whereas today it fails 30% of the time, and by this time next year it should fail just 3% of the time, as shown in the two charts below.¹⁵

Source: ARK Investment Management LLC, 2026, based on data from Datacurve.ai as of July 30, 2026. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security or cryptocurrency. Past performance is not indicative of future results.

Source: ARK Investment Management LLC, 2026, based on data from Datacurve.ai as of July 30, 2026. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security or cryptocurrency. Past performance is not indicative of future results.

A third measure of frontier model performance is more subtle. At the optimum point on the performance-cost curve, what is the tradeoff in dollars for increased or decreased performance on the benchmark? Prior to the debut of test-time compute with OpenAI’s o1 model in September 2024, the user could not test the tradeoff easily. Now, the slope of these models—representing the translation of an incremental percent of budget into an error rate reduction—seems to be steepening. The models are not only becoming more performant at the top-end and less expensive for a given level of performance, they are also becoming more sensitive to user performance requirements. The slope is improving roughly 2-fold per year as measured by the percent of budget dollars required to increase performance by one percentage point, as shown below.¹⁶

Source: ARK Investment Management LLC, 2026, based on data from Datacurve.ai as of July 30, 2026. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security or cryptocurrency. Past performance is not indicative of future results.

The combination of these factors forecasts extreme improvements in performance, both anticipated and realized, across the entirety of the DeepSWE benchmark. From left to right across the chart below, each square represents the realized decline in cost-per-performance either experienced (in the long rectangles) or modeled. Across benchmark scores of 5% to 70%, costs should decline ~99.97% on average over the next year. At higher levels of performance, the cost declines should be steeper, unlocking levels of performance previously impossible.¹⁷

Source: ARK Investment Management LLC, 2026, based on data from Datacurve.ai as of July 30, 2026. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security or cryptocurrency. Past performance is not indicative of future results.

A natural question arises: could cost declines become too steep, creating too much intelligence? According to the evidence thus far, within the context of our forecasts, the answer is no. While AI systems can achieve remarkable feats, they remain woefully inadequate in many knowledge-work areas. In embodied and biological AI applications, they are rudimentary. At every capability threshold crossed so far, advances in model performance have scaled revenue at an unprecedented rate, a trend that we expect to continue. Already, frontier model companies may have crossed $100 billion in run rate revenue, as shown below. By 2030, their revenues could reach into the trillions.¹⁸

Source: ARK Investment Management LLC, 2026, based on data from Anthropic, Series H announcement (May 28, 2026); TickerTrends, private-market ARR tracking (July 2026); Axios/Funda revenue estimates (July 2026); CNBC, OpenAI July revenue update (July 2026); Reuters, Bloomberg, The Information and Bloomberg Law historical ARR disclosures. 2030 frontier-model revenue reflects ARK’s moderate uptake forecast and is not company guidance as of July 30, 2026. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security or cryptocurrency. Past performance is not indicative of future results. Forecasts are inherently limited and cannot be relied upon.

Regardless of frontier revenue trajectories, end-users and enterprises will have access to remarkable levels of intelligence at very low prices per task. Given that resource, the competitive dynamics between and among enterprises should become more fluid and fierce—motivating the winners to continue spending at the frontier of intelligence.¹⁹