Forces · May 2026

Where exponentials collide with atoms. All at once.

The same near-vertical line is showing up in capability, usage, revenue, and token demand. Each curve eventually has to pass through compute, memory, networking, power, and land. That is where physical walls, sheer scale, and shifted economics appear.

01 PhysicsScaling

Compute Density

+66% TSMC N2 wafer cost vs N3

Curve: Model capability and training demand
Then: Cost per transistor halved every two years from 1965 until the curve broke around 5nm in 2020.
Now: Chiplets and packaging carry today's load. The path past EUV is unsettled: free-electron-laser sources, X-ray lithography, no consensus.

02 PhysicsScaling

Memory

92× HBM bandwidth per GPU · 2014 to 2026

Curve: Model size, context, and inference volume
Then: DRAM sat flat on the motherboard for decades. Bandwidth was bounded by the perimeter pins of a 2D die.
Now: HBM stacks 8 to 16 DRAM dies vertically. Rubin will carry 22 TB/s per GPU in H2 2026.

03 PhysicsScaling

Advanced Packaging

858 mm² Single-die reticle ceiling

Curve: Compute density
Then: Plastic substrates expand six times faster than silicon. At reticle scale the package warps under heat.
Now: The monolithic die era is over. Embedded silicon bridges carry today's parts. Glass substrates are the roadmap fix, not yet in production.

04 PhysicsScalingEconomic

Networking & Optics

200 Gb/s PAM4 copper signaling ceiling

Curve: Cluster scale and token throughput
Then: Copper Ethernet has carried the data center since 1973. Insertion loss climbs with the square root of frequency.
Now: Co-packaged optics ship in 2026. Goldman models optical content growing from $15B to $154B by 2028.

05 PhysicsScaling

Power & Cooling

30–120× Rack power growth, 1990s to Rubin Ultra

Curve: Tokens per watt and rack density
Then: Air-cooled racks at 48 volts held from the 1990s through the mid-2010s. Densities crept up. Architecture held.
Now: Power architecture rebuilt at 800V DC. Direct liquid cooling at the cabinet. Rubin Ultra racks hit 600 kW.

06 ScalingEconomic

Data Center Scale

1.2 GW Stargate Abilene, one campus

Curve: AI demand converting into physical capacity
Then: AWS's entire Northern Virginia region ran 500 MW across dozens of buildings in 2015.
Now: One Abilene campus draws 2.4× that whole AWS region. Multi-gigawatt sites under construction.

07 ScalingEconomic

Inference Distribution

>50% CoreWeave inference share of compute, Q1 2026

Curve: Usage shifting from training to inference
Then: Training was the workload from 2012 through 2024. Inference was the afterthought you paid for at the end.
Now: CoreWeave runs majority-inference. Google processes 16 billion tokens per minute. AWS AI annualizes above $15 billion.

08 ScalingEconomic

Application Migration

10× Anthropic ARR, 2023 to May 2026

Curve: Lab revenue and application demand
Then: Going from zero to $100M wasn't hard. Inference in 2023–2024 was cheap because there wasn't much of it.
Now: $1B (2024), $9B (2025), $30B (Apr), $44B (May). The next 10× requires compute that does not exist yet.

The Apollo program cost $323 billion in today's dollars and landed twelve men on the moon over eleven years. The hyperscalers are spending more than twice that this year alone. They are creating intelligence from sand.