Loading...
Loading...
H200 clearance to Alibaba, Tencent, ByteDance and JD.com adds an entirely new demand vector that wasn't in any sell-side model six months ago. Morgan Stanley already revised hyperscaler capex to $805B in 2026 and $1.116T in 2027 — and that was before China came back online. The model doesn't lie: consensus FY27 EPS for Nvidia is still anchored to a world where China is zero. Strip away the noise and you're looking at 15-25% earnings upside that nobody has refreshed their spreadsheet for yet. When deep value and structural growth converge, that's where the asymmetric returns live.
Two pre-IPO giants worth a combined $2.75T are about to become the largest non-hyperscaler concentration of AI compute demand on (and above) the planet. Orbital compute solves the three constraints strangling terrestrial AI buildouts: power, cooling, and land. The chain reaction lights up a supply chain almost no public investor is positioned for — rad-hardened silicon, space-grade memory, optical interconnect, and launch capacity. The market is pricing this like it's a linear story. It's not. If even 10% of the announced orbital compute roadmap materialises by 2030, a handful of small-cap suppliers become household names.
Gold's smashed through $3,000 and central banks are still buying with both hands. That's not a hedge against inflation, it's a vote of no confidence in the US Treasury market. The producers are minting cash at these prices but the juniors (the explorers with real ounces in the ground) are trading at 80-90% discounts to NAV. This is the fattest pitch I've seen in twenty years in the sector. The disconnect cannot last.
An H100 rack draws 40kW. A B200 rack pushes past 130kW. Air cooling physically cannot handle these densities, which is why liquid cooling is going from optional to mandatory across every hyperscaler buildout. The companies making precision cooling, heat exchangers, and coolant distribution units are sitting on multi-year order backlogs while trading at single-digit forward earnings multiples in some cases. The market has stopped looking at this. That's exactly why we should.
Anthropic's declaration that shares transferred without board approval are void has detonated under the SPV industry. The chain reaction extends far beyond one company: every secondary platform peddling pre-IPO access to retail investors is now staring at a counterparty risk problem they cannot price. Combine that with preference stack waterfalls (TSG's 18% compounding BrewDog deal is the cautionary tale), zero secondary liquidity, and post-IPO lock-ups, and the expected value math on private market exposure looks nothing like the marketing decks. The model doesn't lie. Most retail investors in private markets are paying venture capital prices for sub-venture capital outcomes.
The memory market has fundamentally broken from its historical pattern, and Wall Street hasn't updated its models. Micron trades at a 6x forward P/E after a 124% YTD run because analysts still cling to the cyclical playbook. HBM4 is the third leg of a structural demand shift (compute, networking, and now memory bandwidth) that hyperscalers cannot build around. With $625B in 2026 capex commitments and AMD's Helios racks shipping with 31TB of HBM4 per system, the supply-demand math doesn't reconcile until 2028 at the earliest. This is where the asymmetry lives.