Summary

Stargate’s $500B is first and foremost a phased, four-year investment and construction plan—not cash already in the bank. The first-stage target is $100B, with SoftBank handling financing, OpenAI operations, Oracle implementation, and MGX financial investment; NVIDIA, Microsoft, and Arm are named as technology partners. Around 10 to 20 data-center buildings have already broken ground in Texas. “Money is the most direct yardstick for everything in the adult world” （钱是成年人世界衡量所有事情最直接的标准）, so the project’s first question is not the vision but “who is putting up the money?”
The disclosed funding still leaves a huge gap to $500B. Zhuang Minghao cited reports putting the figure at roughly $45B-$49B, but the itemized commitments—$19B each from SoftBank and OpenAI, and $7B each from Oracle and MGX—add up to $52B. 马斯克 immediately questioned, “They don’t have the money.” SoftBank’s roughly $200B and MGX’s roughly $100B in AUM cannot all be committed to a single project, so more equity or debt funding may be needed.
Debt financing would quickly turn the grand vision into a cash-flow test. If $200B of the eventual $500B came from debt, 4% interest alone would cost $8B a year; even with perpetual debt and no principal repayment, early-stage construction revenue may not cover the interest. Equity investors also need an exit path: whether the new company will go public and when shares can be monetized remain unanswered.
The real center of the 14-minute launch was Oracle, not OpenAI or SoftBank. Trump spoke for roughly half the event, Larry Ellison spoke for nearly 4 minutes in total, 孙正义 for about 2 minutes, and Sam Altman for about 1 minute. Ellison also extended the data-center story to detecting cancer 10 to 15 years early and designing personalized vaccines. Zhuang Minghao’s impression was that Oracle occupied the “C-spot,” with its long-standing government relationships and execution capabilities mattering more than its marquee AI status.
NVIDIA remains the most direct beneficiary, but low-cost models are attacking the foundation of its valuation. Zhuang Minghao estimates that roughly 75% of its revenue comes from data centers, with gaming accounting for about 10%; as long as the playbook of “spend aggressively, build infrastructure aggressively, build data centers, buy chips” holds, NVIDIA keeps collecting revenue. The real danger is that DeepSeek and other Chinese models are said to achieve comparable results at 1/50 or even 1/100 the cost of their U.S. peers, challenging the core narrative that more investment produces better models.
Supply-chain opportunities depend on incremental CapEx actually materializing, not on the $500B headline itself. Six large technology companies spent more than $200B on AI-related CapEx in 2024 and plan roughly $330B in 2025; adding Stargate’s first-year $100B would bring the total to $430B, close to a year-on-year doubling. But power, construction, copper cable, and interface infrastructure assets have already rallied over the past year. Investors still need to judge how much of the upside is priced in and whether giants such as Apple, Meta, and Amazon will join.
The project also reshuffles the power relationship among OpenAI, Microsoft, and 马斯克. Microsoft is listed as a technology partner but says it does not understand the specifics of implementation. Its cloud, API, and revenue-sharing ties with OpenAI, which run through 2030 or until OpenAI reaches $100B in profit, may or may not cover Stargate—a huge question mark. 马斯克 is absent from the core plan. Zhuang Minghao believes Sam has gained a higher-level identity from which to keep competing with 马斯克, with both sides competing “from Trump’s shoulders.”
For China, the key is not copying America’s capital-intensive path but deciding whether to follow suit. Stargate could resemble Apollo or the Star Wars program, using massive investment to force a rival onto the same track. But the progress of DeepSeek, Kimi K1.5, 智谱, 豆包, MiniMax, and Chinese robotics companies points to another possibility. Central and local governments, major companies, and the “six little tigers” of startups all face choices across compute, models, applications, and vertical use cases. The ultimate determinant of victory “is something no one can answer.”

Deep dive

1. Tech Giants Move to Trump’s Front Row; Regulatory Winds May Shift

Zhuang Minghao read a political signal from the seating at the inauguration: the heads of 马斯克, Google, Meta, Apple, and Amazon were positioned ahead of many government officials. Microsoft and NVIDIA were the only absentees among the “Magnificent Seven”; NVIDIA’s 黄仁勋 was in China visiting customers and attending the company’s annual meeting.
His view is that over the past 1 to 2 years, not only leading companies but also Silicon Valley funds and investors have broadly backed Trump. Those resources will seek “so-called returns” over the next 4 years, potentially pushing antitrust and big-tech regulation in a more permissive direction.
He described foreign policy more cautiously as “somewhat more conciliatory”: China remains Trump’s strongest rival, but his post-inauguration messaging included seeking a visit to China within 100 days. Zhuang Minghao explicitly acknowledged that he is not an international-politics researcher and that this was only an inference from the surface signals of policy.

2. Stargate Starts as a $500B Four-Year Plan

The project is described as a cumulative $500B investment over the next 4 years, with an initial $100B target focused mainly on AI data centers and related infrastructure. Zhuang Minghao understands it as a joint venture to be formed in the future.
The division of labor is fairly clear: 孙正义 will serve as chairman, SoftBank will handle financing, Sam Altman and OpenAI will run operations, and Oracle will manage concrete implementation. MGX is the AI investment fund of a UAE sovereign wealth fund, while NVIDIA, Microsoft, and Arm are the announced technology partners.
Oracle has begun building the first data centers in Texas, reportedly involving around 10 or 20 large buildings. Zhuang Minghao retained the uncertainty around the figure but emphasized that the project is not entirely confined to a launch event and PowerPoint slides.

3. The 14-Minute Press Conference Made Oracle the Real Protagonist

Based on Zhuang Minghao’s timing of the video, Trump spoke for roughly 7 minutes in total. Of the remaining time, Larry Ellison spoke twice for nearly 4 minutes combined, 孙正义 for about 2 minutes, and Sam Altman for roughly 1.5 minutes. “Based on the distribution of speaking time, the real protagonist was Oracle.”
Trump’s core narrative was jobs, U.S. AI leadership, and “Make America Great Again.” Ellison first thanked the new administration, then stressed that several companies had already been working together and had started construction. Sam, by contrast, spent half his time giving thanks and half repeating OpenAI’s familiar vision, “without much information value.”
Before 孙正义 spoke, staff added a riser because of his height. During his remarks, he spent most of the time turned toward Trump and repeatedly thanked him for the opportunity. Zhuang Minghao found the posture “somewhat obsequious,” while also distinguishing SoftBank’s earlier commitment to invest $100B in the U.S. from Stargate’s own funding.
The language on stage reflected the different positions of power: Trump supplied the national narrative, Oracle the implementation narrative, SoftBank the financing commitment, and OpenAI the AI brand. Oracle, which actually controls construction progress and government trust, was therefore closer to the “C-spot” than the more closely watched Sam.

4. The Cancer Narrative Extends Compute Investment into Public Health

After Trump called on Ellison for a second time, the project’s vision jumped from data centers to medicine: AI could one day use blood tests to identify an individual’s likely cancer 10 to 15 years in advance, then design a personalized vaccine for prevention and treatment.
Zhuang Minghao called it “a gigantic promise built around cancer.” It extends the vision of the data-center and infrastructure project into medicine and oncology.

5. $500B Is Astronomical, but Fits AI’s Step-Change in Investment Scale

Zhuang Minghao drew three annual-investment thresholds: hundreds of millions of dollars for challengers such as Midjourney, Stability AI, and early OpenAI; $1B to $10B for major players such as OpenAI, Anthropic, and xAI; and $10B or more for the large-tech tier.
The figures he cited put the combined AI-related CapEx of 6 large technology companies above $200B in 2024; Amazon, Microsoft, and Meta alone can each exceed $10B in a single quarter. OpenAI’s roughly $2B in revenue and $5B in losses in 2024 also show that spending tens of billions is simply the cost of staying in the race.
Even larger figures had appeared earlier: Sam had proposed a $5T-$7T AI investment vision; SoftBank pledged $100B in U.S. investment; Microsoft planned to spend $80B in 2025; and GAIIP, launched by BlackRock, Microsoft, and MGX, set an initial target of $30B and an overall target of $100B.
His conclusion was not that $500B would be easy to achieve, but that the number was “relatively rational” at the scale of the industry. If the goal is genuinely to rebuild large-scale AI infrastructure, the required capital may indeed be measured in the hundreds of billions. The problem is that after the scale was announced, the funding sources still had to be verified one by one.

6. First-Round Commitments Leave a Huge Gap to $500B

The reports Zhuang Minghao saw used different figures for confirmed first-round funding, at roughly $45B or $49B. But adding SoftBank’s and OpenAI’s $19B each to Oracle’s and MGX’s $7B each produces $52B. Even if fully delivered, that would cover only about one-tenth of the total target.
On the day the news broke, 马斯克 said directly, “They don’t have the money,” adding that based on his information SoftBank could invest no more than $10B. That may have been rival-to-rival criticism, but it also identified the plan’s most verifiable weakness: an announced commitment is not the same thing as cash that can already be drawn.
The plan may later raise more equity or debt. On the equity side, the only clearly visible large financial investor is MGX, and a single $7B check is already astronomical for any fund. Who else will join, what returns they will demand, and whether the company will ultimately go public all remain unclear.
SoftBank’s roughly $200B and MGX’s roughly $100B in AUM mean that neither can “put all the money in its fund into this one project.” Since 孙正义 is responsible for financing, he will still need to find other sources of capital.

7. Debt Costs Turn the Grand Vision into a Cash-Flow Test

Zhuang Minghao’s question is how the project could raise large-scale debt before it has revenue. Borrowing may become easier only after the project reaches a certain scale, generates revenue, and enters a self-reinforcing development cycle.
His stress test is straightforward: if $300B of the $500B comes from equity and $200B from debt, 4% interest would mean $8B in annual interest alone. “Even if this were perpetual debt and you never had to repay principal,” whether early-stage revenue could withstand the burden would remain “a huge question mark.”
That is where “money is the most direct yardstick for everything in the adult world” lands: equity investors need an exit path, while lenders need clarity on maturity, interest rates, and whether the project’s future revenue and cash flow can service the interest.

8. SoftBank, OpenAI, Oracle, and MGX Each Have Their Own Agenda

SoftBank invested roughly $500M in a round valuing OpenAI at $157B, then bought about $1.5B of employees’ old shares at year-end, bringing its cumulative investment to roughly $2B. Alongside its $100B U.S. investment commitment to Trump, it is both a shareholder and the party responsible for organizing new funding.
SoftBank still owns roughly 90% of Arm, which may also be one of its largest U.S. operating assets. Even if data centers remain dominated by NVIDIA, no company wants to be “held hostage” by a single chip architecture. Zhuang Minghao believes Arm may use the project to secure a share of the AI-chip market.
OpenAI is facing pressure on market share—one third-party report he cited says its share fell by the low-to-mid teens over the past year, with roughly two-thirds flowing to Anthropic’s Claude and one-third to Google—while also escalating its competition with 马斯克. Becoming Stargate’s operator gives Sam a higher-level role.
Oracle brings government relationships and implementation capabilities. Zhuang Minghao traced its history of serving U.S. intelligence agencies, supporting Trump, and once being included in a proposed structure for TikTok’s U.S. business. MGX has invested in G42 and infrastructure, and participated in GAIIP and SoftBank’s Vision Fund, potentially making it an important bridge between 孙正义 and Middle Eastern capital.

9. NVIDIA Continues to Benefit as Long as the “Stack Compute” Narrative Holds

Zhuang Minghao estimates that about 75% of NVIDIA’s revenue comes from data centers, while its once-core gaming business accounts for only about 10%. As long as the industry continues building data centers, buying GPUs, and stacking data and compute, 黄仁勋 only needs to “make money comfortably.”
The risk is not a single quarter of orders but a change in methodology. If good models can be achieved without such massive compute and capital, NVIDIA’s core growth narrative will be damaged. “Until this narrative structure is overturned, the stock remains buyable; once it comes under question and attack, it becomes dangerous.”
DeepSeek and other Chinese reasoning and open-source models represent that challenge. What shocked U.S. companies was not only the performance but the possibility that these firms achieved comparable results at 1/50 or even 1/100 the cost. If the low-cost path becomes mainstream, the economics of $500B in infrastructure investment will have to be recalculated.
Zhuang Minghao therefore compares Stargate with the Apollo or Star Wars programs: by making the vision sufficiently large and pushing investment sufficiently high, it may pull rivals into a capital-intensive race. But whether China should follow suit has no obvious answer.

10. Ownership and Potential Expansion Expose a Governance Vacuum

Zhuang Minghao repeatedly asked, “Who owns this company?” Even if the shareholders are OpenAI, Oracle, and SoftBank, the project still needs to clarify who holds the largest stake and who will be the long-term biggest beneficiary. In his view, it is unrealistic for a Japanese company to become the biggest beneficiary of a U.S. national AI project.
He also asked where Trump’s own interests lie. Trump has only 4 years in office but is a shrewd businessman who has “done business all his life”; Zhuang Minghao believes he must have a financial interest in the project, though who will benefit most remains unanswered.
The project could also expand beyond data centers. Trump once proposed that “America” hold 50% of TikTok’s U.S. business, but which entity would hold it on behalf of “America” still needs to be specified. Zhuang Minghao offered one purely speculative possibility: Stargate could take it on, similar to the previous proposal involving Microsoft, Oracle, and Walmart.

11. Incremental CapEx Will Revalue the Supply Chain and Force Absentees to Take a Position

The 6 large companies plan roughly $330B in AI CapEx in 2025, up by more than 40% from 2024. Adding Stargate’s first-year $100B would bring the total to about $430B, close to twice the 2024 level. Power plants, construction, copper cable, interface equipment, and the data-center supply chain could all face incremental demand.
But Zhuang Minghao did not equate that directly with a buy signal: infrastructure stocks have already rallied over the past year. “Who benefits when this money is spent” still has to be considered alongside valuation. Investors need to identify who will actually convert the spending into orders and determine how much of the $500B is merely a long-dated promise.
Giants such as Apple, Meta, and Amazon that are not directly involved also need to decide whether to follow suit. They already have enormous budgets; whether Stargate will prompt them to increase spending remains an open question.
Microsoft’s position is particularly awkward. It is a technology partner, but its CEO said he did not understand the specific implementation and only confirmed that the company’s $80B plan would continue. Whether its exclusive cloud relationship with OpenAI, Azure API arrangement, and revenue-sharing ties running through 2030 or until OpenAI reaches $100B in profit cover the new company’s revenue remains “a huge question mark.”

12. The “$2T Question” Ultimately Comes Down to Whether China Follows Suit

Zhuang Minghao extends Sequoia’s earlier “$600B question for AI”: since ChatGPT launched, industry investment may already have exceeded $1T; once Stargate and the following year’s plans are included, “$2T is already unstoppable.” Technology may have an investment phase, but someone still has to answer when that money will generate returns.
“Sovereign AI” is becoming a shared narrative. Countries are building their own data centers, models, and application ecosystems, while governments are becoming core customers for startups. Citing Palantir as the best-performing stock in the S&P 500, he believes an ecosystem of funds and companies serving governments has already formed. China and the U.S. are in some ways “on the same frequency” in using government guidance to drive AI investment.
China’s national strategy does not need to copy America’s. The progress of DeepSeek, Kimi K1.5, 智谱, 豆包, MiniMax, and Chinese embodied-intelligence companies over the past 6 months at least shows that software models and robotics may develop cost structures below those of the U.S. capital-intensive route. Beijing must weigh whether to follow suit, while local governments must choose among compute, binding themselves to core companies, direct investment, and other approaches.
For startups such as the “six little tigers,” the question is harsher: giants can keep spending, but startups must place their bets on technology, products, government relationships, or verticals such as healthcare, then sustain employees, investors, and the next funding round through staged delivery. Zhuang Minghao’s honest conclusion is that even if you asked 李飞飞, Sam Altman, or Ilya, no one knows whether the ultimate determinant of victory will be the very thing being bet on today. “No one can answer.”

Language Models are "Modelling The World" [Nicholas Carlini]Emergency Pod: Reinforcement Learning Works! Reflecting on Chinese Models DeepSeek-R1 and Kimi k1.5