ARK • Disrupt

This week, AI scare stories swamped the news cycle, catalyzed by a former Anthropic employee who claimed that AI posed a meaningful extinction risk to humanity. After only three to four months at the company, his announcement comes amid staggering performance improvements in AI systems—cracking decades old mathematical conundrums,¹ extending cost decline curves that track to 99.99% annualized,² and demonstrating new world-class computer-use capability.³ As a result, investors and policymakers are highly sensitive to the implications. 

When someone suggests that the implications of new technologies are bad, investors and policymakers want to hear what he or she has to say. As it happens, the Anthropic employee has no special knowledge, his talking points indistinguishable from the ungrounded assertions shared on internet message boards for years and years. Nonetheless, he did at least 10 national media interviews in less than 72 hours.

Distilled, his argument runs as follows:

1. AI will become extraordinarily capable, extraordinarily quickly.

2. Unless properly aligned with human interests, AI will pursue goals without regard for human life.

3. In pursuing its goals, AI will have to acquire resources and power.

4. Humans will compete with the AI system for resources and try to prevent it from accruing power.

5. Therefore, the AI system will have an incentive to eradicate humanity.

Sounds like science fiction? Well, we think it is.

Nonetheless, many people are paying close attention, so we would like to deconstruct the logic carefully.

True, AI tools are becoming extraordinarily powerful extraordinarily quickly. From that statement forward, ARK’s views on the matter diverge from the hysterical.  

A useful way to frame the stacked assertions is to understand that we already depend heavily upon synthetic super-intelligent systems. In a narrow sense, those systems are misaligned with humanity: in pursuing their goals, they do need resources and power. Humans contend with these systems daily and try to prevent them from accruing power. In this case, the “super-intelligences” are corporations, legal bodies driven by 10s or even 100s of thousands of human brains, legally obligated not to operate in the interest of humanity but to seek profit.

The future that the AI safety scaremongers fear so greatly is already here! Yet, it has not resulted in immiseration or extinction, but in remarkable human uplift.

Why don’t corporations trample humans at scale to reach profit goals? Corporate actions are scaffolded by laws, institutions, and property rights that have evolved over centuries to coordinate positive sum behavior. Corporations act in this way, not for the sake of humanity, but because the pathway to accruing and retaining a higher volume of resources requires following the rules. Companies that do not follow the rules have a difficult time accruing resources, particularly at scale. Sure, organized crime is powerful, but it thrives only in locations with relatively weak institutional scaffolding. Indeed, the most powerful organized crime conglomerates are tiny compared to the world’s largest businesses.

In our view, AI systems will be scaffolded in a similar way. The pathway to power lies in the rules of law, economics, and markets.

Tangibly, we can see this pathway evolving in AI systems. The frontier labs need compute, and their proxies—primarily, infrastructure-as-a-service—are racing to build it out. When humans try to stop them, the infrastructure-as-a service providers don’t devise plans to eradicate the local population. Instead, they pony up resources—taxes, community centers, power price protections and guarantees—to cover the perceived downsides of their infrastructure buildout. If the price rises too high, they find less expensive areas in which to build. They also are creating ways to provide more efficient and convenient infrastructure to coexist with humans, orbital data centers an important example.

Rhetoric about human extinction is inflammatory, so it catches the eye. But it is ungrounded and demonstrates a profound misunderstanding of complex systems. In this way, it rhymes with the ongoing conjectures that AI is going to eradicate tens of millions of jobs, when in reality the rise of AI already has created roughly 1 million new jobs in the US alone, as shown below.

*ARK-independent share-only signal. Bars use 12-observation smoothed total employment × (12-observation smoothed share—baseline share pooled over the 12 months ending Nov 2022); macro total-employment growth is excluded from the plotted value. Positive share shifts rise and negative share shifts fall below zero. Missing source months are not imputed; trailing windows use the available observations. Source: ARK Investment Management LLC, 2026, based on data from the Bureau of Labor Statistics as of September 14, 2026. BLS/Census Current Population Survey public-use microdata, final person weights. October 2025 CPS was not collected. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security.

*ARK-independent share-only signal. Bars use 12-observation smoothed total employment × (12-observation smoothed share—baseline share pooled over the 12 months ending Nov 2022); macro total-employment growth is excluded from the plotted value. Positive share shifts rise and negative share shifts fall below zero. Missing source months are not imputed; trailing windows use the available observations. Source: ARK Investment Management LLC, 2026, based on data from the Bureau of Labor Statistics as of September 14, 2026. BLS Current Employment Statistics, seasonally adjusted payroll employment. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security.

Even if the rhetoric is wrong, the political impact could become real. The most likely byproduct of any legislation would cement the competitive position of the current AI leaders and slow the rate of overall progress, bad for humanity but probably good for OpenAI, Anthropic, Google and xAI who can bear the fixed cost of conforming to a regulatory regime while continuing to deliver revolutionary products.

Potentially, some level of regulatory pressure would destroy growth entirely, though that outcome is not the most likely: AI tools are so useful that billions of people globally already are using them. That people rely upon them protects them somewhat from the most draconian regulatory outcomes.

Ultimately, ARK’s research suggests that these tools will prove to be incredible instruments for human uplift and wealth creation. Today they are solving decades-old insoluble math problems. Tomorrow, they could be driving material science breakthroughs, drug development breakthroughs, manufacturing engineering breakthroughs, and cures for disease worth in dollars billions,10s of billions, even hundreds of billions and trillions over the course of this decade.

What an awful shame it would be if rhetorical science fiction did anything to derail those gains!

Last week, Meta launched Muse,⁴ a personal agent that books travel, clears inboxes, fills forms, and completes purchases. In our view, Muse is the first consumer agent with all the features and functionality necessary to scale.

Designed for the household, Muse can plan a night out, shop for a vacation outfit, lower a monthly bill, and check a home camera. The app also carries a “Feed” tab with a personalized stream from a user's connected accounts and from the web, a capability that matters more than it appears: with an interface users already know, the feed gives people a reason to open Muse on days they are not actively engaging with their agents. AI chief Alexandr Wang said that most people will use the free tier, with the $20 and $100 plans limited to the heaviest compute users.

Meta is optimizing first for adoption, not for revenue per user. Mark Zuckerberg has indicated that initially Meta will take a "very small percentage" of the transactions that its agents complete. ARK's research suggests that AI agents could facilitate roughly $9 trillion in payments, approaching 25% share of e-commerce, in 2030.⁵ With lead generation take rates in the mid-single digits, that revenue pool could top $200 billion, equal to Meta's current advertising revenue per year. Moreover, a “free” agent clearing and scraping transaction fees should be worth more than a “paid” agent answering questions. 

Distribution makes the model work. Last quarter, Meta reported ~3.6 billion daily active human users across its family of apps, describing WhatsApp as the leading app for its AI engagement. SpaceXAI also runs Grok Bot’s persistent agents on its own cloud infrastructure but charges users a minimum $200 per month. Instinct, the text-and-call assistant that raised $350 million at a $2.5 billion valuation in August,⁶ remains invite-only. Both are real agent experiences, but neither is likely to reach even a billion people this year. 

We expect Meta to take the lead in consumer agents because it knows how to build consumer products, can get them in front of users without paying to acquire them, and can deliver a compelling experience at a cost few other companies can match.

A perpetual contract (“perp”) is a derivative native to crypto that gives traders synthetic directional exposure to an underlying asset. Purely synthetic, perps are not backed by the underlying assets: they offer bets on the direction that an asset will move, typically with a matching engine pairing one trade against the opposite trade. Perps use a price feed called an “oracle” to track the underlying asset's spot price on exchanges and a mechanism called “funding rates” to anchor the trading price to that oracle price. One side of the trade pays the other to pull the price back to equilibrium with the spot price on an exchange: if the price moves above the spot market price, longs pay shorts a small fee to encourage more short positions and close the spread.

Onchain perps have taken off for two reasons. First, retail traders find the user experience (UX) of perps simple compared to dealing with rolling at expiration dates. Second, perps live onchain and offer the following benefits: 24/7 access, global collateral, and liquidity sourced from a global, region-agnostic base. Historically, centralized Asian exchanges like Binance and OKX have dominated perps trading,⁷ but recently Hyperliquid has emerged as a real competitor, leveraging its onchain footprint and an extremely efficient exchange architecture⁸ to deliver superior execution.

Only recently have US markets tried to onshore perps. Kalshi and Coinbase were the first movers⁹, launching native regulated perps, Coinbase structuring its offering as a 50-yr dated futures contract. The growth has been exceptional. In response, incumbent exchanges ICE and CME have petitioned the CTFC (Commodity Futures Trading Commission) to slow progress with more rigorous requirements.¹⁰ Meanwhile, Hyperliquid has created a US policy advocacy group, the Hyperliquid Policy Committee (HPC), led by prominent crypto lawyer Jake Chervinsky, to serve as its advocate in the US.¹¹ Adding to the drama, at the recent White House crypto summit, Trump named Hyperliquid as a platform likely to be onshored in the US. 

Whoever prevails, we believe perps are here to stay and are likely to be offered broadly across US platforms.

Last week, Amazon announced a partnership with OpenAI that will bring ChatGPT advertising inventory into Amazon’s Demand Side Platform (DSP),¹² a demand-side software platform for advertisers to buy and manage digital advertising inventory across publishers. The deal comes shortly after OpenAI’s advertising business hit a ~$1 billion annualized revenue run rate (ARR) and The Trade Desk announced plans to cut its workforce by 15%.¹³ In other words, Amazon has strengthened its position in the DSP market by adding premium inventory and is pressuring standalone competitors with cut-rate fees.

The economics of Amazon’s DSP are compelling. Relative to The Trade Desk’s implied take rate of ~22% and Google’s 15% for display inventory purchased through DV360, Amazon is charging 0% for its own media and ~1% for select streaming inventory.¹⁴ Amazon can support much lower fees because its DSP drives demand for its own advertising inventory, data products, and Amazon Web Services (AWS). Over the past several years, Amazon has expanded its DSP beyond Amazon.com by adding premium inventory from Netflix, Disney, ESPN, Roku, and others. Products like AWS Clean Rooms also protect the privacy of retailers’ first-party data on Amazon, addressing a historical source of friction in advertising with their large retail competitor.

For years, The Trade Desk’s “Switzerland” pitch differentiated it from walled gardens that owned the inventory they sold. That advantage seems to be diminishing as Amazon combines commerce data, premium inventory, measurement, and cloud infrastructure in one platform. Amazon’s ability to cross-sell to companies already using AWS is a significant competitive advantage as marketers link advertising spend to business outcomes.

¹ OpenAI. 2026. “On the Navier–Stokes Millennium Prize Problem.” OpenAI.
² Winton, B. 2026. “AI Costs Are Plummeting.” ARK Investment Management LLC.
³ OpenAI. 2026. “GPT-6 Astra: A New Generation of Intelligence.” OpenAI.
⁴ Meta. 2026. “Introducing Muse: The World’s First Personal AI Agent Built for Everyone.”
⁵ Grous, N. and A. Kim. 2024. “Generative AI: A New Consumer Operating System.” ARK Investment Management LLC.
⁶ Ropek, L. 2026. “Viral AI startup Instinct has raised $350M at a $2.5B valuation.” TechCrunch.
⁷ Blockworks Research. 2026. “Crypto Exchanges.”
⁸ Hyperliquid Docs. 2026. “Order Book.”
⁹ Hamilton, J. and N. De. 2026. “U.S. CFTC opens crypto 'perp' door with first approvals at Kalshi, Coinbase.” CoinDesk. See also Park, D. and V. Chawla. 2-25. “Coinbase launches CFTC-regulated perpetual futures for US traders amid growing crypto regulatory clarity.” The Block.
¹⁰ Braun, A. 2026. “CME, ICE push U.S. regulators to scrutinize Hyperliquid over manipulation risks.” CoinDesk.
¹¹ Roberts, J.J. 2026. “Hyperliquid launches DeFi-focused policy shop led by prominent crypto lawyer Jake Chervinsky.” Fortune.
¹² Palmer, A. 2026. “Amazon gives OpenAI’s ad business a boost, letting its advertisers into ChatGPT.” NCBC.
¹³ Ostwall, T. 2026. “Is The Trade Desk Running Out of Room to Turn Around?” Ad Week.
¹⁴ Hercher, J. 2026. “The Trade Desk Faces Headwinds As Investors Reconsider The Thesis Of Objective Indie Ad Tech.” Ad Exchanger. Hisaio, S. 2020. “How our display buying platforms share revenue with publishers.” Google Ad Manager. MacLean, K. 2025. “Simplicity and scale—how the Amazon DSP is redefining advertiser value and success.” Amazon Ads.