SpaceX's Record IPO and the Age of "AI as Weapon and Heavy Industry"

Hello, this is Ryuta Hamamoto from TIMEWELL.

On June 12, 2026, SpaceX went public on Nasdaq. Ticker SPCX, offering price $135 per share. Reports put the valuation at about $1.77 trillion and called it the largest IPO in history, surpassing Saudi Aramco in 2019.¹ This is a loss-making company. Its 2025 net loss was about $4.9 billion, and its accumulated deficit stands near $41 billion.² And still the market set the price at $1.77 trillion.

I did not read this news as a mere "rocket company IPO." I see it as a miniature of the era that is coming. In one line: AI is a weapon, and it is heavy industry. At the entrance to that era stands SpaceX, as a symbol. In this article I trace SpaceX from its prospectus through Starlink, Starship, and the merger with xAI/Grok, against the primary sources, and build up that thesis. Technology, geopolitics, and money get tangled here, so I take it step by step. At the end, I consider what Japanese companies should take from it.

Loss-making, yet the largest ever — an IPO pricing "the future"

First, the numbers. SpaceX offered about 556 million shares at $135, raising roughly $75 billion in the base offering, up to $86 billion including the over-allotment.¹ The first-day close was about $161; on June 16 it peaked in the $225 range, then fell over three sessions to settle around $150 by late June. It did drop from the high, but it never fell below the $135 offering price. "The stock fell after the IPO" is, more precisely, "a pullback after an opening pop."

The interior is what interests me. In the S-1, 2025 revenue was about $18.7 billion — and about 60% of it, $11.4 billion, came from Starlink.² Less a launch company now than a satellite-connectivity company. Meanwhile Starship's R&D burns around $3 billion a year and is spelled out over many pages as a top-tier risk that generates no revenue. Elon Musk retains roughly 80% of the voting power through super-voting Class B stock even after listing, and this "dependence on a single person" is likewise flagged as a risk.

So why does a loss-making company command $1.77 trillion? My reading is this. The market is betting not on current profit but on Starlink's near-monopoly growth as communications infrastructure, and the future value of a structure that binds launch, connectivity, space, and AI in one company. The stock is trying to price "dominance ten years out," not "today's income statement." If that reading is right, SpaceX's valuation is also a vote from the market on where AI-era industry is heading.

If you want to take stock of how far you have woven generative AI and its surrounding infrastructure into your strategy, measure your position first with the AI literacy check; the later analysis will feel more like it is about your own company.

Three failures, and a company that nearly went bankrupt

When I tell SpaceX's story, this is the part that draws me most. It is the world's largest space company now, but it once nearly went under completely.

Falcon 1, the early workhorse, was — in SpaceX's own definition — a two-stage, liquid-fueled small-lift launch vehicle that operated from 2006 to 2009. The prospectus states that "with the first successful launch of Falcon 1 in 2008, we became the first private company to successfully launch a liquid-fueled rocket to Earth's orbit."³ Which is to say, the launches before that did not reach orbit. The often-told version — three consecutive failures, with the fourth flight riding on the last money Musk personally had left — comes from his own later recollections, and I could not corroborate it against a primary record. What is documented is that in the same year, 2008, NASA awarded fixed-price contracts to Orbital ATK and SpaceX for ISS cargo resupply — 20 missions initially, across the two contracts combined — with SpaceX's task orders initially valued at about $1.6 billion. That contract is what let the company survive.³

I put this story first because SpaceX's decision-making habit is concentrated in it. Fail repeatedly, then bet the remaining cash on the next attempt. This "bet the company" posture runs through the later expansions of Starship and Starlink. Even as a public company, it keeps burning $3 billion a year on Starship. You cannot explain that behavior with the financial discipline of a normal listed company — but this company has stacked up enough story and track record to make it tolerable.

Starlink — a "communications moat," quietly completed

The first card of the vertical stack is Starlink. Per SpaceX's S-1 filed with the SEC, as of March 31, 2026 the company had approximately 9,600 Starlink broadband and mobile satellites in Low-Earth Orbit — about 75% of all active maneuverable satellites in orbit.⁴ Over 3,000 of them were deployed in 2025 alone, the constellation supports over 700 Tbps of cumulative downlink capacity, and it serves approximately 10.3 million subscribers across 164 countries, territories, and other markets. The US FCC has already authorized 12,000 first-generation and 15,000 second-generation satellites — 27,000 in total — and on an application basis the figure reaches about 42,000.⁵ The "up to 40,000 satellites" idea refers to this application ceiling. It is a slot they hold, not a confirmed build plan — worth keeping in mind. There is also talk of extending the network to the Moon and beyond, but that is still a future sketch, not a fixed plan.

The world realized Starlink was not "just fast internet" in Ukraine. Since 2022, Ukrainian forces have been widely reported to use Starlink as a backbone for command communications, drone operations, and artillery correction.⁶ Even when ground infrastructure is destroyed, the network overhead survives. The US Department of Defense was also reported to have signed a contract in 2023 to cover its operating costs. Both rest on reporting; I could not confirm them against a primary source. You often see it put as "Starlink's support prevented Ukraine's defeat," and here I write carefully. The contribution was large — that is a fact. But asserting the causal claim of "prevented defeat" has no clear expert consensus behind it. That is an evaluation, not a verified fact.

At the same time, the reality that a private company can, on its own judgment, sway the shape of a war was laid bare in Ukraine. In September 2022, when Ukraine sought expanded coverage to attack the Russian fleet near Crimea, Musk was reported to have declined to "activate" it, citing potential conflict with US sanctions — again, something I could not confirm against a primary source.⁶ As for the government-facing Starshield, SpaceX itself states in its prospectus that it is "a secure satellite network designed specifically for national security applications," focused on three core mission areas: Earth observation, global secure communications, and hosted payloads.⁶ The same filing says its government revenue comes from long-term Starshield contracts, without naming the counterparties. Communications infrastructure has become economic security itself. Here is the seed of the "AI is a weapon" argument I make later.

Starship, and the idea of "designing the experience"

The second card is Starship. A fully reusable super-heavy rocket, it has run integrated test flights through the first half of 2026 and pulled off the feat of catching the booster in midair with the launch-tower arms (nicknamed Mechazilla) several times.

The thing I personally find interesting about Starship is the "Earth to Earth" concept Musk announced in 2017.⁷ Travel across the Earth by rocket. Los Angeles to Tokyo in about 32 minutes; New York to Shanghai in 39. Most long-haul trips in under 30 minutes. That is the source of the "fly Japan to Los Angeles in a bit over 30 minutes" story. But to be honest, this is a 2017 concept, and as of 2026 not even a test flight has happened. G-loads during acceleration and deceleration, offshore launch pads, noise, regulation, safety. The unsolved problems are legion. That it is an unrealized concept is a premise you cannot drop.

Even so, I value this concept because they are trying to redesign the experience of travel itself. Not make airplanes faster, but connect cities with rockets. The starting point sits outside the extension of existing industry. This "work backward from the experience" posture — separate from feasibility — helps explain why people and money gather around this company. And do not forget: what Starship really delivers is not passengers but the destruction of launch cost. Move mass to orbit cheaply and in bulk, and Starlink, and the space-infrastructure ideas I discuss next, suddenly become plausible.

Data centers in space — but "it's cold, so it cools fast" is a myth

Here is where AI and space intersect. In recent years, putting AI data centers in orbit has been discussed seriously — startups like Starcloud, and Google's research project (Project Suncatcher), are examples. GPU clusters on the ground eat power, cooling water, and land. So, the thinking goes, do it in space.

But let me state one misconception clearly. "Space is cold, so you can cool GPUs quickly" is physically wrong. In a vacuum there is no convection or conduction; the only way to shed heat is infrared radiation. Radiative heat rejection is the slowest form of heat transfer and scales with the fourth power of temperature, so it needs enormous radiators. In fact, industry press identifies cooling as one of the biggest technical barriers for orbital data centers. The real advantage companies cite is "no cooling water needed (saving ground water resources)," not "faster cooling." This is a point to nail down precisely, without being carried away by the vision.

For solar power, the advantage is real. Orbital solar panels see no atmospheric absorption and are almost unaffected by weather or day-night cycles, so Google says the energy yield can be "up to about 8x" in the right orbit. The "7 to 10 times" figure refers to this ratio of annual energy output. The Moon comes up too. Permanently shadowed craters at the lunar poles sit at an average of about 40 K, extremely cold, and there is a concept that this could sharply cut the cooling burden for quantum computers; Musk himself has referenced "a quantum computer on the Moon" on X. But this is about the polar craters, not "the far side" in general, and no funded, real project is confirmed yet. There is also a 2.6-second round-trip signal delay between Earth and Moon. It is a concept with imagination — but for now it is a concept.

The reason I think this cluster of ideas matters is separate from timing. AI's constraint has moved from algorithms to physics — power, cooling, land, launch. And space is being floated as the place to break through that physical constraint. That very structure is evidence that AI has become a heavy industry.

The completed form of vertical integration — SpaceX × xAI × Tesla

Then in 2026, a decisive piece snapped into this picture: the merger of SpaceX and xAI. In February 2026, the two announced a combination valued at about $1.25 trillion, reported as the largest merger in history.⁸ With it, the AI model Grok, the social network X, and xAI's massive data centers were bound to the SpaceX side. Compute (Grok, Colossus), connectivity (Starlink), space access (SpaceX), and — linked through capital and procurement — energy and manufacturing (Tesla). Under one umbrella of capital, almost all of the "physics" of the AI era is now assembled. This is the completed form of vertical integration.

The model side is moving too. Grok 4.5, released on July 8, 2026, is a cheap model tilted toward coding and agent use, priced at $2 per million input tokens and $6 per million output — reportedly more than 60% below top-tier models.⁹ Musk called it "Opus-class." You also hear "equivalent to Opus 4.8" among users, but independent benchmarks place it around fourth, slightly short of the very top — that is the neutral read as of July 2026. "Cheapness is the main weapon" is the accurate framing. I do not assert the ranking, since it changes constantly. Read this alongside how Kimi closed in on world-class with open models, and the map of model competition comes into three dimensions.

But vertical integration has a shadow. In mid-July 2026, Grok Build — a developer tool — was reported to have uploaded entire repositories, including secrets like API keys and passwords, to the cloud, beyond what the task required.¹⁰ The privacy toggle reportedly did not stop the upload itself. After disclosure, uploads were halted, and Musk said the data would be deleted. Some call it "customer code sent arbitrarily," but the reporting treats it as a bug or design flaw, not an intentional act; I see no basis to call it "malice" at this point either. Even so, the danger of one entity holding compute, data, model, and customers is well illustrated by this incident. Around the same time, there were reports that old code from Cursor (developed by Anysphere), then in acquisition talks, had contaminated Grok's training data and given it an edge on a benchmark. Vertical integration is powerful, but it is also a breeding ground for conflicts of interest and concentrated power. I note both sides, calmly.

AI is heavy industry — the bottleneck has moved to electricity

Here I assemble the backbone of this article. First, from the "heavy industry" side.

That AI's constraint has moved from compute resources to physical infrastructure is now a shared recognition among the principals. In 2024 Musk said, "A year ago the shortage was chips; next it's transformers, and then it will be electricity," and in 2025 he said the rapid expansion of compute could bring a power-generation shortfall from mid-to-late 2026.¹¹ NVIDIA's Jensen Huang frames the economics of AI data centers as "revenue is roughly tokens per watt × available gigawatts," placing power as the primary lever of AI revenue. Zuckerberg and Altman say, in unison, "the constraint from here is power." The International Energy Agency (IEA) projects that data centers' global electricity consumption will roughly double to about 945 TWh by 2030.¹²

This no longer wears the face of a software industry. Large data centers, power plants, transformers, copper, cooling, and capital expenditure on the order of hundreds of billions of dollars. AI is taking on the character of "heavy industry," the same as steel, chemicals, and shipbuilding. It is turning into an industry where whoever holds capital and physics wins. That is exactly why SpaceX's vertical integration matters. Not only compute, but the power to run it (Tesla), the connectivity to carry it (Starlink), and even space as the place to shed heat — all captured in one company. If that is not "AI as heavy industry," what is? For fairness: "heavy industry" is the framing of analysts and the press, not a stock phrase of Musk's. What he actually talks about is the industrial structure in which the bottleneck chains toward electricity.

AI is a weapon — between rockets, nukes, and the US-China divide

The other face is "weapon." I write this carefully, but plainly.

Musk himself has spoken of AI's danger from early on. In 2014 at MIT he said, "With artificial intelligence we are summoning the demon," and that it could become "our biggest existential threat";¹³ in 2018 he said "AI is far more dangerous than nuclear weapons," calling the lack of oversight "insane."¹⁴ In 2023 he signed an open letter calling for a six-month pause on training giant AI. The irony is that this same Musk now leads one of the world's largest complexes of AI, space, and communications.

And SpaceX's business is contiguous with weapons technology to begin with. Rockets (launch vehicles) and ballistic missiles share core technologies — propulsion, staging, guidance. Put a nuclear warhead and re-entry vehicle on a ballistic missile and it becomes a nuclear missile. That is exactly why the Missile Technology Control Regime (MTCR) exists. "Put a warhead on a rocket and it's a nuclear missile" is not sensationalism but a classic dual-use fact. To be fair, though, the grid fins and propulsive landing for reuse are mainly about cutting cost, and are not themselves a weapons program. I want to avoid the shortcut of "reusable rocket therefore weapon."

In this context, China's move cannot be overlooked. On July 10, 2026, state-owned CASC recovered the first stage of its Long March 10B on its debut flight, catching it with a net on a sea platform.¹⁵ It became the second country after the US to recover an orbital-class booster. But what was achieved was "recovery"; re-flying the same vehicle, "reuse," is still to come, planned within the year. The method also differs from SpaceX's leg landing, using hooks and a net. On the AI side, too, China's presence is growing. Alibaba's Qwen went from 3.5 to 3.7 through 2026, and in July a 2.4-trillion-parameter Qwen3.8-Max was announced as a preview. DeepSeek has released new models as well. Against the backdrop of US GPU export controls, Chinese players are trying to reduce dependence on NVIDIA.

Space access, satellite communications, AI models. All three become instruments of national security as they are. In Ukraine, communications infrastructure swayed the course of a war; rocket technology is contiguous with missiles; and the founder himself calls AI "more dangerous than nukes." Saying AI is a weapon is by no means an exaggeration. And as long as the US and China collide head-on in this domain, the position of a single private company, SpaceX, becomes, like it or not, a variable in geopolitics. That is why I believe this company should be watched closely going forward, including from a military standpoint.

Hamamoto's take — how should Japan fight "AI as heavy industry"?

From here, this is my view. How should Japanese companies receive SpaceX's IPO and vertical integration? I write in three parts.

First. If AI has become heavy industry, Japan should, if anything, have a path to win. If AI's main battleground is shifting from clever algorithms to "physics" — power, cooling, materials, manufacturing, space — that is precisely where Japan was strong. Power generation, electrical equipment, precision manufacturing, materials, and heavy industry. Ironically, a Japan that fell behind at the entrance of generative AI may find its cards more usable the more AI turns heavy and large-scale. I see this as opportunity, not pessimism. I wrote before about how few Japanese firms appear on NVIDIA's list of AI-native companies; if the industry's center of gravity moves to physics, the roster can still change.

Second. Do not treat the "concentration risk" behind vertical integration as someone else's problem. A structure in which one entity holds compute, data, model, customers, and communications is powerful — and, as the Grok Build incident showed, also a breeding ground for accidents and concentrated power. Entrust your confidential data, the core of your operations, and even your communications infrastructure entirely to a specific overseas platform, and your business is left to the discretion of that entity. The structure is the same as what Ukraine experienced over Crimea. That is exactly why the option to run frontier models under your own control, and a design that preserves data sovereignty, carries more strategic weight than ever. Our insistence on domestic-server enterprise AI with ZEROCK comes down to this single point.

Third. Smaller companies, of all people, should imitate the "work backward from the experience" posture. SpaceX attracts people and money not because its finances are a model student's, but because it is trying to fundamentally redesign the experience of travel, of communication. Japan's AI-adoption discussions still tilt toward "which tool do we adopt" and "how do we cut cost." Instead: can you fundamentally redesign your customers' experience and your own operations with AI? Before it is a technology question, it is a question of vision.

To be honest, I do not offer unqualified praise for SpaceX. Dependence on a single founder, the scale of the losses, concentrated power, and proximity to weapons technology. I see the dangers as dangers. And yet the outline this company reflects — that "AI is a weapon, and heavy industry" — will, like it or not, define the next ten years of industry. Do we watch rockets tear the sky as a flashy piece of news from a far country, and leave it there? Or do we connect it to our own decisions? That gap will tell over the next decade. That is how I see it.

If you want to think together about how to redesign your business and operations on the premise of AI, you can reach us through our AI-adoption support program WARP or an individual consultation. Translate the "heavy-industrialization of AI," advancing behind the flashy space news, into your own strategy. That is what I most wanted to convey in this article.

---

References

This article is composed by cross-checking SpaceX's official IPO announcements and SEC filings (the S-1), FCC authorization documents, news reports, and statements by the principals. Figures and dates are based on each source; for reported figures such as valuation and share price, and for forward-looking concepts, that status is noted in the text. Statements by Elon Musk and other CEOs are referenced as close as possible to primary reporting and originals.

Footnotes

1. SpaceX, "Space Exploration Technologies Corp. Announces Pricing of Initial Public Offering" (June 11, 2026) — 555,555,555 shares of Class A common stock at $135.00 per share, expected to begin trading on the Nasdaq Global Select Market and Nasdaq Texas on June 12, 2026 under the ticker "SPCX," with a 30-day underwriter option for up to 83,333,333 additional shares. The $75 billion base offering and the ~$86.2 billion maximum including the over-allotment are arithmetic from those share counts and that price. https://content.spacex.com/cms-assets/FINAL_Documents%20and%20Updates/SpaceX_PricingAnnouncement.pdf / Note: the "$1.77 trillion valuation" and "largest IPO in history" are not in this official announcement; they are reported figures, and the Form S-1 (filed May 20, 2026) is a preliminary registration statement that leaves post-offering shares outstanding blank, so the valuation cannot be derived from primary sources. The body text marks them as reported. ↩ ↩²

2. SpaceX Form S-1 (SEC EDGAR, CIK 1181412). https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm / financial breakdown: https://www.satellitetoday.com/finance/2026/05/20/spacexs-ipo-filing-gives-first-look-into-companys-financials/ ↩ ↩²

3. Falcon 1 reaching orbit: SpaceX Form S-1 (SEC EDGAR, CIK 1181412, filed May 20, 2026), verbatim "With the first successful launch of Falcon 1 in 2008, we became the first private company to successfully launch a liquid-fueled rocket to Earth's orbit." and the filing's own definition of the term, "'Falcon 1' refers to our two-stage, liquid-fueled small-lift launch vehicle that operated from 2006 to 2009." https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm / The NASA resupply contract value: NASA Office of Inspector General, "Audit of Commercial Resupply Services to the International Space Station" (Report No. IG-18-016, April 26, 2018), verbatim "In 2008, while development efforts were still underway, NASA awarded fixed-price contracts with task orders initially valued at $1.9 billion and $1.6 billion to Orbital ATK and SpaceX, respectively, for 20 cargo resupply missions to the ISS through 2016." Those 20 missions are the initial total across both companies' contracts, not SpaceX's own count; the same report adds that later task orders brought CRS-1 to 31 missions. https://oig.nasa.gov/docs/IG-18-016.pdf / The three consecutive failures and the "last remaining money on the fourth flight" account come from Musk's own later recollections and could not be verified against a primary record. The CNBC article previously cited here now returns HTTP 404. ↩ ↩²

4. Starlink constellation scale. SpaceX Form S-1 (SEC EDGAR, CIK 1181412, filed May 20, 2026), verbatim: "As of March 31, 2026, we had approximately 9,600 Starlink broadband and mobile satellites in Low-Earth Orbit, operating the world's most advanced broadband constellation providing internet connectivity to approximately 10.3 million Starlink Subscribers across 164 countries, territories, and other markets." On share, elsewhere in the same filing: "approximately 9,600 Starlink broadband and mobile satellites in Low-Earth Orbit, which accounted for approximately 75% of all active maneuverable satellites in orbit as of March 31, 2026." On capacity: "Our current constellations of approximately 9,600 Starlink broadband and mobile satellites, including over 3,000 satellites deployed in 2025, support over 700 Tbps of cumulative downlink capacity." https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm ↩

5. FCC authorization/application for the second-generation Starlink. https://docs.fcc.gov/public/attachments/DA-26-36A1.pdf / https://www.scientificamerican.com/article/spacexs-starlink-constellation-could-swell-by-30-000-more-satellites/ ↩

6. Starshield is characterized per SpaceX Form S-1 (SEC EDGAR, CIK 1181412): "Separately, we operate Starshield, a secure satellite network designed specifically for national security applications," "Starshield is focused on three core mission areas: Earth observation, global secure communications, and hosted payloads," and, on revenue, "We generate government revenue via long term contracts for Starshield." Elsewhere the same filing describes it as designed for "United States Government customers and national security applications." https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm / Note: the phrase "Department of Defense" does not appear in the filing at all, and the National Reconnaissance Office (NRO) appears once, listed among the customers of SpaceX's launch services alongside NASA and others — it is not named as a Starshield counterparty. / By contrast, Starlink's battlefield use in Ukraine, the US Department of Defense covering its operating costs, and the September 2022 decision on expanded coverage near Crimea all rest on media reporting that we could not confirm against a primary source ("Ukraine" appears zero times in the S-1). ↩ ↩² ↩³

7. Starship "Earth to Earth" concept (IAC 2017). https://techcrunch.com/2017/09/28/spacex-wants-to-use-its-giant-bfr-rocket-for-earth-trips-too/ ↩

8. Reporting on the SpaceX–xAI merger (February 2026, ~$1.25T valuation). https://www.cnbc.com/ ↩

9. Grok 4.5 release and benchmark positioning (third-party evaluation such as the Artificial Analysis Intelligence Index). https://artificialanalysis.ai/ ↩

10. The Grok Build repository-upload issue (July 2026). https://www.axios.com/ ↩

11. Elon Musk's "chips → transformers → electricity" bottleneck argument (Bosch Connected World 2024) and the power-shortfall outlook. https://www.cnbc.com/ ↩

12. IEA outlook for data-center electricity consumption. https://www.iea.org/reports/electricity-2024 ↩

13. Elon Musk, "we are summoning the demon" (MIT, October 2014). https://www.washingtonpost.com/news/innovations/wp/2014/10/24/elon-musk-with-artificial-intelligence-we-are-summoning-the-demon/ ↩

14. Elon Musk, "AI is far more dangerous than nuclear weapons" (SXSW, March 2018). https://www.cnbc.com/2018/03/13/elon-musk-at-sxsw-a-i-is-more-dangerous-than-nuclear-weapons.html ↩

15. China's Long March 10B first-stage recovery (CASC, July 10, 2026). https://spacenews.com/ ↩