press → to begin

Three sellers. Three record quarters. Three falling prices.

Who was selling

What the company reported

What the price did

Korean retail leverage

After May's single-stock leveraged ETFs, one in thirty adults held a leveraged account that got margin-called.

SK hynix 000660

Revenue +257%
Operating income +557%

−15.4%

one session, 13 July

Institutional leverage

Situational Awareness LP, ~4× levered.
$45bn → $10bn in days.

CoreWeave CRWV

Revenue +112%
Backlog $104bn

−30%

month to 26 July

Narrative shock

1 July: Meta to resell surplus compute.
MS Momentum TMT −53.5%.

Bloom Energy BE

Revenue +166%
Product revenue +215%

−32%

across July

1

Q2 2026 · Q2 Earnings

Four sentences from the quarter

"Even at that amount, we will still not have enough capacity to meet all the demand we have in 2026… In fact, the demand we already have for 2028 is striking."

Andy Jassy, CEO, Amazon — while raising 2026 capex from $200bn to $220bn. AWS backlog: $496bn.

"We're able to deploy capital in such a way that we're getting less than a one-year payback. It's almost like your capital is becoming an item you would otherwise put in COGS, because it's moving so quickly into revenue."

Bret Johnsen, CFO, SpaceX — while Ed Zitron calls the private credit funding these builds "the subprime mortgage crisis of the AI bubble."

"We recently signed an A100 contract that extends into 2029 at an attractive price. As a reminder, this SKU was introduced in 2020."

CoreWeave, Q2 call — while Burry calls writing off AI chips over five to six years "one of the more common frauds of the modern era."

"SpaceX expects to buy more than five gigawatts of Nvidia's Vera Rubin architecture next year — over $200 billion. Nvidia's entire revenue last year was $216 billion."

Jukka Lepikkö on X

3

Q2 2026 · AI Capability Curve

AI builds AI — faster every cycle

The software

Code shipped per Anthropic engineer

0.9×

Q4
2023

1.0×

Q1
2024

0.9×

Q2

0.9×

Q3

1.0×

Q4

1.2×

Q1
2025

1.5×

Q2

1.9×

Q3

2.5×

Q4

5.8×

Q1
2026

8.0×

Q2
2026

●Anthropic Institute — multiple of pre-2025 avg

The iron

Autonomous chip verification, weeks to hours

WAS

weeks

⟶

NOW

hours

40×

faster verification cycles

Cadence "ChipStack" super-agent. NVIDIA's first customer is itself — verifying its own next-gen chips.

●Jensen Huang, GTC Taipei 2026

"We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for."

Q2 2026 · AI Capability Curve

Capabilities evolve faster

METR Task Complexity Timeline

Task length at 80% success rate · model release date

Source: METR (metr.org) · CC-BY · doubling time 196 days · R² 0.98

Key Models in 2025–26

Projected capability frontier by model

Asilo Insight: Doubling time: 7 months → accelerating to ~4 months

7

space / → to advance · V to replay · M to mute

click the slide or press M for sound

Q2 2026 · AI Capability Curve

Improvement in quality means capability unlock

Key Models in 2025–26

Projected capability frontier by model

BEFORE

Onefile at a time

JS???

Debug syntax errors

Test & retest

Search documentation

Hours per feature

Brain overload ↓

NOW — EFFICIENCY

OUTPUT

Fourfiles at once

JSJSHTMLJSON

✓ Almost always works

DOWNLOADS FROM CLAUDE

Server v19JS

Metrics calculator v2JS

Index v9HTML

Position history v2JSON

DEPLOY

scp server_v19.js server:
scp metrics_v2.js server:
scp index_v9.html server:
scp position.json server:
ssh systemctl restart

✓ Live in seconds

8

One threshold, six trajectories · V to replay

Q2 2026 · Drivers of AI

Drivers of AI

1

Jevon's Paradox

Lower costs → higher total spending

25 years of semiconductors:
100× cheaper → 3× more revenue

2

AI Speed Revolution

400× faster than Moore's Law

Moore's

2.8x

AI Progress

1200x

3 years of AI:
1200× cheaper tokens

3

Scaling Laws Hold

More compute = better performance

AIME scores: 27% → 90%
No plateau in sight

4

Token Explosion

Scale differenceFrom single prompts to continuous reasoning loops

Multi-step processingIterate, reflect, and refine outputs

AgentsAutonomous systems on complex tasks

Token Explosion: millions of tokens per conversation

9

Q2 2026 · The Analogy Trap

Familiar shape. Wrong limits.

When forecasting AI we instinctively borrow the shape of past growth — and quietly import its limits. They don't apply here.

The borrowed model

Substitution into a fixed pool

better ROI per ad €

LINEAR TVFACEBOOK

Ad dollars moved from TV to Facebook because the unit economics were better. Growth was bounded by the size of the existing ad market — a finite, known pool.

Familiar shape. Built-in ceiling.

Why this wave breaks the model

Four forces, no ceiling

01

New use cases emerge

Capability is multidimensional. Each new threshold unlocks tasks that did not exist before.

Breaks: "the total market is fixed"

02

Jevons paradox at speed

Tokens 1,200× cheaper than 3 years ago. Total spend rises, not falls.

Breaks: "cheaper unit → lower revenue"

03

Scaling laws hold

More compute → predictably better models. No plateau in sight.

Breaks: "technology eventually plateaus"

04

Token explosion

Prompts → reasoning loops → agents. Per-task usage explodes by orders of magnitude.

Breaks: "per-task usage is stable"

10

Contents

Three sellers. Three record quarters. Three falling prices.

Four sentences from the quarter

AI builds AI — faster every cycle

Capabilities evolve faster

METR Task Complexity Timeline

Key Models in 2025–26

Improvement in quality means capability unlock

Key Models in 2025–26

Drivers of AI

Familiar shape. Wrong limits.