Michael Hochberg

NEW: Sunrun to turn rooftop solar systems into “mini data centers,” offering to pay homeowners hundreds per month to support AI compute.

TSMC June 2026 & 2Q26 - June broke seasonality; revenue +6% M/M and +57.4% Y/Y to $14.0B, a new ATH; June declined M/M in each of the last 4 years (avg -11%) and this is the first June M/M growth since 2021 - hitting the $40.2B high end guidance required +4.9% M/M and TSMC delivered +5.8% - 4th consecutive ATH, 5th in 6 months of 2026; incremental +$769M over May breaks the deceleration trend we flagged in May (Jan +6.0%, Mar +2.4%, Apr +0.5%, May +1.1%, now +5.8%)

Compute probably won’t become one uniform market. It may end up segmenting into different tiers, with frontier workloads paying for premium infrastructure while lower value workloads gravitate toward cheaper models and cheaper power. Not every token is equal and not every workload needs the newest chips, lowest latency, and five nines power. Frontier training, high value inference, autonomous systems, drug discovery, financial applications etc will probably keep paying up because the cost of being slower or worse can be way higher than the cost of the compute. But a huge amount of demand will be much more price sensitive. Batch jobs, basic enterprise inference, lower urgency workloads. Those can move to cheaper chips, cheaper regions, (dare I say) interruptible power, and less premium infrastructure. So compute may end up looking less like one giant commodity market and more like different classes of service. Some workloads pay first class prices, a lot of the rest goes with cheaper options.

Just a few problems with McKibben’s claims

We imported more from Taiwan than China as tariffs and AI trends collide. (1/6)🧵