Two trillion-dollar IPOs. One compute lease.

Start with the gap in the valuation itself. The $322 billion AI projection is Goldman's; Morningstar, reading the same filing, puts fair value near $780 billion — under half the ask. When the lead underwriter and an independent desk differ by nearly a trillion dollars on one company, the disagreement is the story, not the number.

H15 was written for this moment, and the filing pulls it in two directions at once. The architecture strengthened: the vertical-integration thesis is no longer inferred from scattered deals but set out in a public prospectus — launch, Starlink, orbital and terrestrial compute, a chip fab with Tesla and Intel, and the xAI absorbed in February. On paper, one owner of the whole stack. The commercial proof weakened in the same pages. The flagship Colossus 1 is leased to a competitor — xAI moved its own training to the newer Colossus 2 and rented the older site out rather than run it — and the AI unit lost $6.4 billion last year. And the lease is not the locked $40 billion windfall the headline rate implies: Musk clarified it runs six months, not four years, with either side able to exit on 90 days' notice, the short term set at SpaceX's own request to keep its options open. The blueprint says commanding. The income statement, and a landlord hedging its own lease, say not yet. The call stays Both Ways — the two readings just sit further apart than ever.

Then the merger talk, which is where discipline matters most. Wedbush's Dan Ives put the odds of a SpaceX–Tesla merger above 80% within a year; Kalshi traders price the same event near 33%. One of those is a headline, the other is money. What is not speculation is the wiring already disclosed: Tesla's $2 billion stake in SpaceX, the shared Texas chip fab, the xAI shares that converted into SpaceX holdings. The merger would formalise a thing the filings show already half-built. Read it that way and it sharpens H15; read it as the 80% number and you've been sold the forecast as a fact.

The clearest read on the power thesis this week wasn't a price. It was a schedule. More than 60 percent of the US data-centre capacity planned for 2027 is not yet under construction, and the named hold-ups are power, permits, and supply chain — not chips. The constraint stopped being a forecast about where the bottleneck would land. It is a description of where it already sits.

The firms with capital are responding the way the thesis predicts: they have stopped being customers of the grid and started building it. Google paid $4.75 billion for Intersect, a wind and solar developer, and is putting on-site generation behind a Texas data centre rather than wait in the utility queue. Policy is moving the same way — Ireland, where data centres already draw a fifth of national electricity, now tells new facilities to bring their own power, and the EU is drafting efficiency standards as demand threatens to double by 2030.

This thread has a precedent worth naming, and a divergence worth drawing. Signal 002 made the same call on the US grid: PJM's wholesale capacity price had roughly doubled on data-centre demand, but the household pass-through was small, 1.5 to 5 percent, and the real exposure was narrative, not economic. The difference now is structural. The US is meeting the demand shock by adding supply across many sources at once — gas, coal, nuclear restarts, and private generation built behind the meter — which gives the system slack. Australia is adding it largely through one channel, the renewables-and-storage transition, on a grid with thin domestic gas to fall back on. Same load, less room to absorb it.

Western Sydney is where that tightness becomes concrete. Western Sydney Airport, the city's first new international airport in over a century, opens late this year, and the surrounding Aerotropolis has pulled private investment from under $10 billion to over $21 billion in fourteen months, with more than 650,000 square metres of data-centre development now in planning. Real jobs and real capital, converging on one corner of the grid. That same corner is where the network operator has flagged the limit: Transgrid's own assessment warns that demand growth around the airport precinct could exceed the capacity of the Macarthur transmission point as early as 2026/27, potentially requiring load shedding, or controlled outages, if upgrades lag the build. The economic case and the grid constraint are landing in the same postcode at the same time. That is the cliff, and it sits in the operator's own filings, not in any forecast we are supplying. H7 strengthens not because a bill number is settled, but because the gap between demand and the grid's ability to carry it is now this specific, and this near.

Signal 003 flagged the top Chinese and US coding models sitting within a single point on SWE-Bench Verified. This week the cluster produced a leader, and it wasn't American. MiniMax shipped M3, built for long-context agentic coding, and reporting put it ahead of GPT-5.5 and Gemini 3.1 Pro on SWE-Bench Pro — the harder variant. One benchmark, one harness, the usual caution on decimals. But the gap that was closing from a distance is now, on at least one credible test, briefly inverted. H6 strengthens on the continuation, not the single score.

The same week, the lab leading this race asked everyone to slow down. Anthropic — fresh off a $965 billion filing built on exactly that lead — publicly urged the field to consider pausing development of the most powerful systems, citing the approach toward recursive self-improvement. Take the safety argument at face value or read the competitive subtext; either way the signal is the same. The people closest to the frontier think the slope is steepening, and they are saying so out loud while the commercial machine accelerates underneath them.

Neither moves H4. A model topping a coding benchmark is not a system running end-to-end research; a call for caution is not a capability. The autonomous-researcher thesis holds Unchanged. But both point the same way underneath: the rate of gain is rising, and the threshold H4 watches for is getting nearer, not further.

The earned side strengthened on hard numbers. Anthropic's $47 billion run-rate is now in an IPO filing, not a private mark. Around it, the proxies reported in audited results: HPE up roughly 40 percent year-on-year on AI-server demand, targets pulled forward; Palo Alto raising forecasts on AI-security adoption. Not projections — results. But the strength has an expiry date worth watching. Anthropic's curve is steep and young, riding an Opus launch that landed just as enterprises started testing intelligence layers at scale. Whether that converts to loyalty or churns when the trials end is untested. The direction is not in doubt. The staying power is.

And the first note of buyer's fatigue arrived this week — from the buyers, not the builders. Commonwealth Bank and Coles were reported reassessing AI rollouts as token pricing turned cheap experiments into real line items. That is the exact language the bear case waits for: optimisation, cost discipline, ROI scrutiny. But H2 is a thesis about a hyperscaler cutting capex, and a bank questioning its AI bill is not Amazon guiding down. So it gets logged, not acted on. H2 stays Weakened, with a watch flag — the discipline being to catch the turn early without mistaking one week of grumbling for it.