ARK Invest
ARK InvestSep 1
Crypto

AI Is Outgrowing Our Own Forecasts | Big Ideas 2026 Mid-Year Review

23 min video4 key momentsWatch original
⚡
TL;DR

AI token inference demand has grown 49x in 18 months, forcing chip spending to $800B annually — far exceeding earlier forecasts — while Anthropic and OpenAI are scaling revenue faster than any SaaS company in history.

Key Insights

1

56 trillion tokens weekly — AI token inference on OpenRouter grew from 7 trillion tokens per week in early 2025 to 56 trillion by mid-2026 — a 7x increase in just seven months, after growing 25x in all of 2025.

2

$800B vs $580B forecast — Annual spending on AI chips jumped from an expected $580B to over $800B in 2026 as hyperscalers raised capex guidance based on actual AI ROI from deployed models.

3

$9B to $47B in six months — Anthropic scaled from $9B ARR at end of 2025 to $47B by mid-2026, matching Salesforce's total ARR in under six months and suggesting the true figure is even higher.

4

Tripled to 180 minutes — AI agent capability tripled in the first three months of 2026, with Claude and Mythic models now handling 3-hour tasks autonomously — up from 12-minute tasks at the start of 2025.

5

$10B added in five months — OpenAI added $10B in ARR in five months post-ChatGPT Work launch, reaching $30B by mid-2026, with July alone generating more ARR than all of Q2 combined.

Want this for every new video ARK Invest posts? Brevyd summarizes each upload automatically, the morning it drops.

Deep Dive

Token Demand Exploding Exponentially

Joseph opens by walking through the most startling number: token inference demand on OpenRouter, the unified API platform, grew 25x across all of 2025 to hit 7 trillion tokens per week. That chart looked exponential when published in Big Ideas 2026, but it's been completely eclipsed by what's happened in the first seven months of 2026 — another 7x growth to over 56 trillion tokens weekly. The drivers are twofold: more efficient models making AI accessible to use cases that were previously unaffordable, and breakthrough reasoning capabilities that unlock higher-end applications. This demand isn't theoretical — it's actual consumption by enterprises and consumers finding immediate value in token-by-token improvements.

Chip Spending Inflecting Beyond Forecasts

Infrastructure spend tells the story of confidence. In the decade before ChatGPT, data center chip investments grew just 5% annually, barely faster than inflation. After ChatGPT launched, that inflected to 2.5x the historical rate. When ARK published Big Ideas 2026 in January, industry consensus pegged 2026 chip spending at $580B — a plateau suggesting growth was slowing. But hyperscalers saw actual ROI materializing, raised their capex guidance, and now that forecast sits at $800B for the year. Google Cloud and AWS are posting accelerating revenue growth as a direct result of these investments, validating the spend. Andy Jasse's comment that AWS could become a trillion-dollar annual business — doubling or tripling from its current trajectory — signals how vast the opportunity still looks internally.

Power Becomes the Binding Constraint

With hundreds of billions flowing into chips, the question shifts: where does the electricity come from? Daniel runs through distributed energy progress. On nuclear, X Energy IPO'd in 2026 as the largest public offering for a nuclear company in history, while multiple advanced reactor startups achieved criticality — a self-sustaining fission reaction. The regulatory tailwind is real: NRC permit review times dropped from three years to 18 months thanks to executive orders, and the agency already executed that timeline in March 2026. Solar is emerging faster, particularly tied to orbital data center buildouts. Elon committed to 200 gigawatts of annual capacity by decade's end — five to six times 2025 US output, though only two-thirds of what China deployed last year. The takeaway: more power supply means lower electricity costs, directly benefiting grid-dependent AI infrastructure despite current political nervousness around energy prices.

Consumer Adoption Outpacing the Internet

VKA shifts focus to the consumption side. Consumer adoption of AI chatbots has outpaced internet adoption — a stunning inflection. The monetization puzzle is clear: $20/month subscriptions alone won't fund the infrastructure. The theory centers on two channels. First, agentic commerce — AI shopping assistants that generate leads and transaction fees. Shopify reported AI-driven traffic up 8x year-over-year and AI orders up 13x, with new buyers from AI channels converting 2x higher than other channels. Amazon merged Rufus into Alexa to create Alexa for shopping; 350 million-plus consumers use it, active users are up 2x, and US customers using Alexa for shopping spend 40% more per order. Second, advertising — OpenAI went live with ads on free tiers, while Anthropic took a Super Bowl-level stand against ads, at least initially. The long game mirrors social networks: build trust and habit first, monetize aggressively once adoption matures.

Enterprise Agents Scale Capabilities and Revenue Simultaneously

Joseph quantifies the agent inflection using the meter benchmark — how long a task takes an AI agent versus a human expert. Start of 2025: 12 minutes. End of 2025: 66 minutes. First quarter 2026: over 180 minutes, with Claude and Mythic handling three-hour tasks autonomously. The improvement stems from model reasoning advances but more critically from harness layers like Claude Codebase Work and ChatGPT Work, which let agents integrate with real company data and applications beyond raw reasoning. This capability explosion immediately unlocked new use cases and drove ARR growth. OpenAI was at $20B ARR when Big Ideas 2026 published; by mid-year it hit $30B, with unconfirmed reports of July alone generating more ARR than all of Q2. Anthropic's arc is even more brutal: $9B at end-2025 to $47B mid-2026, matching Salesforce's total ARR. Vertical software startups followed: Cursor grew from $1B to $4B ARR in the same window before SpaceX's $60B acquisition, Harvey tripled ARR, and others scaled similarly.

Takeaways

  • ✓Track quarterly capex guidance from hyperscalers — when they raise targets mid-year like they did in 2026, it signals confirmed ROI on AI infrastructure, not speculation.
  • ✓Monitor token inference growth metrics from OpenRouter and similar platforms weekly, not quarterly — the acceleration has become the primary signal of AI adoption velocity.
  • ✓If shopping feels frictionless through Alexa or similar agentic commerce tools, that's a sign the monetization inflection is real; test it yourself rather than waiting for earnings calls to confirm.
  • ✓Recognize that 180-minute agent autonomy in Q1 2026 means use cases tractable only six months prior are now commodity; plan internal workflows around agents handling three-hour tasks unsupervised.

Key moments

0:46Token inference explodes 7x in six months

“what we've seen in the first seven months of 2026 is continued exponential growth uh growing an additional sevenfold over the course of the year so far uh to in excess of 56 trillion tokens being inferenced per week”

3:00Chip spending leaps to $800B

“what's actually played out over the course of uh of the year so far is that those estimates have been raised uh to continued inflection and now in excess of 800 billion dollars for the year”

10:07Anthropic scales to $47B ARR in six months

“they've grown their AR from 9 billion at the end of 2020 five to roughly 47 billion uh as of the midpoint of 2026”

16:13Agent capability triples in three months

“what we've seen so far in the first um really not even the first seven months of 2025 but the first three months of 2026 excuse me um we've seen this number triple uh roughly a yen to uh over 180 minutes”

You just read one. Brevyd does this for every upload.

Follow ARK Invest and every new video comes back as a summary like this, in your morning briefing. No watching required.