Compute is the asset to own.

The shortage is structural, not a spike

The bottleneck is booked out for years.

The choke point isn't the chip — it's advanced packaging and memory, and both are sold out. TSMC's CoWoS packaging is booked through 2027, with one buyer holding the majority of it. ² All three memory makers are effectively sold out of high-bandwidth memory into 2026–2027, and the pull is so strong that ordinary DRAM contract prices roughly doubled in a single quarter (≈+93% QoQ, Q1 2026), with another large rise projected for Q2. ¹

Demand behind it is not a blip: hyperscaler capital spending is running ~$725B in 2026, up ~77% year-on-year, ³ and it's now bounded by power, not just silicon — one hyperscaler alone disclosed a multi-tens-of-billions backlog it can't build for lack of electricity. ⁵ When the constraint is fabs, packaging lines and grid capacity, it takes years, not quarters, to clear.

Chips don't obsolete like phones

A GPU cascades; it doesn't expire.

Consumer electronics lose their use the moment the next model ships. AI accelerators don't. A chip earns through a value cascade: frontier training in years 1–2, high-end inference in years 3–4, then bulk and batch inference for years after — roughly 5–7 years of productive economic life, well beyond its accounting depreciation. ⁴ A V100 was retired from a major cloud ~7.5 years after launch; a 2018-era T4 still rents today. ⁴ The hardware keeps paying long after the headlines move on.

The honest version

A value floor — not "only up."

We won't oversell it. The price of any single generation does fall as newer parts arrive — used top-end cards have dropped sharply from their peaks, and older ones lose most of their sticker value over a few years. ⁴ The point isn't that a given card only appreciates. It's that compute as a class stays scarce and stays useful: your node holds a resale floor and, more importantly, keeps earning across the cascade while the whole market's supply stays tight. Own the asset that produces income, not the one that just depreciates in a drawer.

Where we cite forward projections (multi-year shortages, trillion-dollar 2027 capex), treat them as industry outlooks, not settled fact. The hard price and supply figures above are the best-sourced.

So owning beats renting

Rent, and you fund someone else's balance sheet.

Every hour you rent cloud compute, you pay down a hyperscaler's ~$725B of hardware — and hand them your data on the way. Own a node and the same scarce asset works for you: it replaces your SaaS, serves the network, and earns while the shortage holds. That's the whole rationale behind the financial model — the best way in is to get the hardware.

Already bought an expensive rig? Bring it. Ship us your pro GPU (Radeon Pro, RTX, …); we verify it and credit up to 90% of its original unit price toward a node, then ship you the hardware. You upgrade into the network instead of letting scarce silicon idle — the value floor, made literal.