METAMORPHOSIS

From Teraflops to Gigawatts

Metamorphosis: From Teraflops to Gigawatts

Martin Chang · July 2026 · Singapore

I woke up in Singapore to a hyperscaler shedding roughly US$190 billion in market value in the pre-market.1

Half a world away in New York, the market was digesting the largest capital expenditure announcement in corporate history: two hundred billion dollars, under one name and one year.2 To give you some perspective, this was enough to fund two-thirds of the entire NASA Apollo program (roughly US$260–320 billion in 2026 dollars, depending on inflation, with NASA’s $25.8 billion nominal cost by CPI or by aerospace-sector indices).3 Most of it primarily went towards data centers, chips, and power.4

The funniest thing was that Amazon wasn’t even the story. It was the cherry on top of the cake that had been baking for two weeks.

Microsoft had opened up the sequence in late January, reporting US$37.5 billion of capex in a single quarter — a US$150 billion annual pace, up 66%. Furthermore, Microsoft’s report signaled to investors that the spend would keep accelerating.5 Its stock sold off.6 Alphabet followed with US$175–185 billion, which was close to double its 2025 spend of US$91 billion. Sold off.7 Meta guided its number to US$115–135 billion, up from US$72 billion.8 Finally, Amazon walked in last, on a Thursday, and placed the largest capital expenditure number in corporate history on the table.9 By the time I got up in Singapore on Friday morning, the verdict was in and clear: all three out of four companies had a combined commitment of US$490–520 billion — and Microsoft hadn’t even announced their full number yet.10 It was only in April, when the answer turned out to be US$190 billion.11 Bringing the four companies’ total to a record of roughly US$700 billion, 70% more than a year that was itself an all-time record.12

The capital expenditure figure forced everyone to question if this was a bubble. But no one is asking the better question. What if this isn’t an overpriced mania, but a mispriced metamorphosis? The largest software and technology companies that we’ve spent the past twenty years loving because they were asset-light and owned nothing but software and networks are being converted, quarter over quarter at a run rate approaching a trillion dollars a year, into infrastructure utilities. Capital intensity that belongs to railroads, debt that belongs to telecoms, and power contracts that could feed nations. An evolution is underway. The same caterpillar that built apps is becoming a butterfly that owns watts.

And the transition is silent. Not because it is hidden but simply because it is happening in the one phase of the current cycle that the market’s traditional instruments were never built to read in real time. Traditional metrics like free cash flow, operating margin, depreciation — none of those are obsolete, and none of them are wrong. But they are harvesting metrics, and this is planting season.

Table of Contents

Each part is meant to stand on its own, though the essay is written as one escalating argument — gigawatts and contracted dollars, section by section — and is best read in order.

00 Introduction [this page] February 2026: the largest capital expenditure announcement in corporate history, and the market sold. What if this isn’t an overpriced mania, but a mispriced metamorphosis?
01 The Right Metric The dollar line compounds at 1.7–1.8x a year until it breaks — on physics, not financing. Three tiers of capacity (aspirational, contracted, energized), nuclear plants restarted with the builders’ own money, and capital intensity that no longer looks like software.
02 The Demand Is Pre-Sold The evidence, stacked in order of contractual weight: Oracle’s US$553 billion in signed obligations, Microsoft’s US$80 billion unserved backlog, AWS monetizing capacity on arrival — and demand graded by who actually stands behind the contract.
03 Why the Market Cannot See It The software-era toolkit reads a utility buildout as value destruction. From PUE arithmetic to scarcity rent: counting, in watts, what the accounts cannot yet see — and why the moat is the energized interconnection.
04 The Casualties, and the Rebuttals Record backlog on one axis, record default insurance on the other. Oracle, OpenAI, the neoclouds, and circular financing — then the two rebuttals that actually threaten this essay, and what would falsify it.
05 What the World Looks Like If This Is Right The fourth great network buildout of the industrial era, built pre-sold by the most profitable companies in history. The instruments always change last.
§ Works Cited

Notes

  1. Amazon reported Q4 2025 after the close on Thursday 5 February 2026. Shares fell about 11% in after-hours trading and closed down roughly 8% on Friday 6 February. Market capitalisation going into the print was approximately US$2.4 trillion, so an 8% move is on the order of US$190 billion. CNBC, “Amazon stock falls 8% on $200 billion spending forecast, earnings miss,” 6 February 2026, cnbc.com. Across the nine-session slide that followed, Amazon shed more than US$450 billion, its worst run since 2006. ↩
  2. https://www.sec.gov/Archives/edgar/data/1018724/000101872426000002/amzn-20251231xex991.htm ↩
  3. Project Apollo cost $25.8 billion in nominal dollars across FY1960–1973 per NASA’s congressional budget submissions (~$28 billion including Project Gemini and the robotic lunar precursor programs). The Planetary Society’s Apollo cost study, which escalates each year’s actual appropriation using the NASA New Start Index — an aerospace-specific deflator that tracks engineering and materials costs rather than consumer prices — puts this at roughly $257 billion in 2020 dollars. Carried forward to 2026 (CPI-U 258.8 → ~330, a 1.28x multiplier; aerospace escalation at 2.5–4% annually gives a similar 1.16–1.27x range), the full program lands at approximately $300–330 billion in 2026 dollars. A simple CPI adjustment from the program’s spending-weighted midpoint (~1966) gives a lower bound near $260 billion. Amazon’s 2026 capex guidance of ~$200 billion therefore equals roughly two-thirds of the entire Apollo program — per year. ↩
  4. Jassy told analysts the 2026 spend would go “predominantly” to AWS — data centres, custom silicon and networking. Q4 2025 earnings call, 5 Feb 2026. ↩
  5. https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q2 ↩
  6. Microsoft fell 7% in extended trading on 28 January 2026 and closed the following session down roughly 10%, erasing about US$357 billion of market value — its worst single day since March 2020. ↩
  7. Alphabet, Q4 and fiscal year 2025 results, 4 February 2026: 2026 capital expenditure guidance of US$175–185 billion against full-year 2025 capex of US$91.4 billion. The stock fell 7.4% on heavy volume. sec.gov ↩
  8. https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Fourth-Quarter-and-Full-Year-2025-Results/default.aspx ↩
  9. 5 February 2026 fell on a Thursday; Amazon reported after the close, so the market reaction landed on Friday 6 February. ↩
  10. Arithmetic: Alphabet US$175–185 billion + Meta US$115–135 billion + Amazon US$200 billion = US$490–520 billion. ↩
  11. Microsoft, fiscal year 2026 third quarter results, 29 April 2026. CFO Amy Hood guided calendar-year 2026 capital expenditure to roughly US$190 billion, up about 61% from 2025 and some US$35 billion above the US$154.6 billion Visible Alpha consensus, of which approximately US$25 billion reflects higher component pricing. microsoft.com ↩
  12. Arithmetic. 2026: Microsoft US$190bn + Alphabet US$180bn (midpoint) + Meta US$125bn (midpoint of the January guide) + Amazon US$200bn = US$695bn. 2025 actuals: Amazon US$131.8bn + Alphabet US$91.4bn + Meta US$72.2bn + Microsoft calendar-2025 of roughly US$118bn (FY25 Q3 US$21.4bn + FY25 Q4 US$24.2bn + FY26 Q1 US$34.9bn + FY26 Q2 US$37.5bn) ≈ US$413bn. Growth ≈ 68%. ↩

This work is dedicated to my mentor, Mr. Robert Ludwig, whose guidance, generosity, and belief in me have shaped this work and my thinking.

Thank you to everyone who has supported and helped me throughout this process. Special thanks to Mr. Arvind Khattar for his invaluable guidance and encouragement.

And to my friends, Martin Lim, Jonathan Quek, and Skye Flecker — thank you for your support, conversations, and friendship along the way.