SpaceX to Raise $75 Billion|Rewire Evening News Update

marsbitDipublikasikan tanggal 2026-03-30Terakhir diperbarui pada 2026-03-30

Abstrak

SpaceX has filed for a $750 billion IPO, targeting a $1.75 trillion valuation—the largest in history, surpassing the combined U.S. IPO totals for 2024 and 2025. The offering, underwritten by Morgan Stanley and Goldman Sachs, is set for June. In AI infrastructure news, Starcloud raised $170M to deploy orbital data centers with H100 GPUs, while Mistral borrowed $830M to build a Paris data center. Meanwhile, the U.S. faces power constraints, with 7.5GW of wind projects frozen by military reviews. Palantir is deepening its U.S. government ties, working with the IRS to build AI tools for audit targeting and a unified data system, raising privacy concerns. In Iran, tensions escalate as Trump administration rhetoric shifts toward resource control, with military movements reported and oil prices surging. Other updates include DeepSeek’s major outage, Dell’s $25B AI infrastructure growth, and the successful IPO of Chinese robotics firm Huayan. Ethereum Foundation also made a large ETH stake, signaling a strategic shift.

Rocket builders are set to swallow the entire IPO market, while algorithm writers help the IRS select audit targets. There's not enough electricity for AI, so someone is sending GPUs into space.

1| SpaceX Aims for $75 Billion IPO, Set to Swallow the Entire Tech Listing Market

SpaceX has filed its prospectus with the SEC, targeting a $75 billion fundraising round at a valuation of $1.75 trillion. This would be the largest IPO in human history—2.5 times the size of Saudi Aramco's $29.4 billion record—and exceeds the total IPO fundraising in the U.S. for 2024 and 2025 combined. Morgan Stanley and Goldman Sachs are underwriting the offering, with a target listing date in June.

After acquiring xAI, SpaceX is now a satellite internet provider (Starlink has 9 million users and profits exceeding $8 billion), a defense contractor, and an AI infrastructure company. Its valuation has surged 38-fold in six years, from $46 billion to $1.75 trillion.

OpenAI and Anthropic are next in line. The $75 billion fundraising scale raises a bigger question than SpaceX itself: how much market liquidity will it absorb, and what valuations will the queued AI unicorns be able to secure?

(Source: Bloomberg / Axios / The Information / Barron's)

2| New Dimensions in the AI Infrastructure Race: Space GPUs, European Debt Financing, and 7.5GW of Frozen Wind Power

Starcloud has completed a $170 million Series A round, reaching an $1.1 billion valuation and becoming Y Combinator's fastest unicorn ever, just 17 months after Demo Day. Using $3 million in seed funding, the company launched its first satellite equipped with H100 GPU, aiming to build an orbital data center network of 88,000 satellites. With Earth's electricity in short supply, someone is starting to send GPUs into space.

Mistral has borrowed $830 million from seven banks to build a data center in Paris, equipped with 13,800 GB300 GPUs, set to go online in Q2. In Europe, AI funding isn't coming from VCs; it's coming from banks.

In the U.S., the Pentagon has frozen military approvals for 30 wind power projects, totaling 7.5 gigawatts. The American Clean Power Association has given the Defense Department until April 8th to respond, or face a lawsuit. While AI companies scramble to build data centers, power approvals are stuck in paperwork.

(Source: TechCrunch / SpaceNews / Reuters / Axios / American Clean Power Association)

3| Palantir Selects IRS Audit Targets, AI Enters the Core of Federal Law Enforcement

Documents obtained by Wired show the IRS paid Palantir $1.8 million last year to improve a custom tool called SNAP, which sifts through legacy systems to select "highest-value" targets for audits, tax debt collection, and potential criminal investigations. The tool extracts key information on contracts, vehicles, and suppliers from unstructured data, replacing over 700 outdated screening systems the IRS has used for decades.

A larger project is the "mega API." Palantir is working with IRS engineers to build a unified data interface layer, aiming to become the "read center for all IRS systems." Democratic senators have already demanded Palantir explain this "searchable super-database" plan, calling it a "surveillance nightmare."

Palantir has received over $180 million from the IRS since 2018 across 26 contracts. The same company that tracks immigrants for ICE, selects audit targets for the IRS, and integrates government data for the DOGE is building the federal government's data hub—a private company with a market cap exceeding $100 billion.

(Source: Wired / Tax Notes / U.S. Senate Finance Committee)

4| Day 31 of the Iran War: Trump Says "Seize the Oil," Marine Corps Already Deployed

Trump has publicly stated he would invade Iran to "seize the oil," calling opponents "fools." The Marine Corps' 31st and 11th Marine Expeditionary Units have arrived in the Middle East, with the 82nd Airborne Division en route. The Washington Post reports the Pentagon is preparing for "weeks of ground operations," including seizing the Kharg Island oil hub. The war narrative has shifted from "retaliation" to "resource control."

The IAEA has confirmed Iran's heavy-water reactor is no longer operational. Oil prices have risen to $116 per barrel, up over 50% since the war began. EIA data shows flow through the Strait of Hormuz has plummeted from a pre-war 20 million barrels per day, forcing Gulf countries to cut production by millions of barrels per day.

Macquarie analysts believe Trump is following a "TACO" timeline (Trump Always Chickens Out), expecting the conflict to end within six weeks—now in its fifth week. The economic cost: approximately $1 billion per day in spending, with over 3,000 deaths. (Continued from yesterday's report)

(Source: Al Jazeera / Fortune / CNN / The Washington Post / Military.com / EIA / IAEA)

Also Worth Knowing ↓

DeepSeek experiences its longest outage since launch, affecting 355 million users for 7 hours. China's most popular AI chatbot suffered its most severe service disruption since the R1 launch in January 2025. The cause of the outage was not disclosed. (Source: Bloomberg / Reuters)

Dell built a $25 billion AI infrastructure business from scratch in two years. The "outdated PC maker," whose stock fell by a third in 2022, reached a record total revenue of $113.5 billion in 2025 and is guiding Wall Street to expect AI server sales of $50 billion next year. Nvidia isn't the only one selling picks and shovels in the AI gold rush. (Source: Fortune)

Huayan Robotics rose 8.24% on its Hong Kong listing debut, with its public offering oversubscribed by 5,000 times. The company listed on the Hong Kong Stock Exchange at an issue price of HK$17 and a market cap of HK$9 billion, securing nearly $100 million in cornerstone investments from Hillhouse, GF Fund, Morgan Stanley, and others. The embodied intelligence capital frenzy is spilling over from the primary to the secondary market. (Source: 36Kr)

Ethereum Foundation makes a one-time stake of $42 million worth of ETH. On-chain data shows this is the Foundation's largest single staking operation to date. The Foundation had long avoided staking, making this strategic shift noteworthy. (Source: CoinDesk / The Block)

Chinese cloud providers achieve scaled profitability for the first time, as AI demand reverses industry losses. Tencent Cloud turned a profit for the first time in 2025, and Kingsoft Cloud reported positive profits for two consecutive quarters after adjustments. Cloud providers' pricing power is increasing in the AI era, potentially ending the history of low-price competition. (Source: 36Kr / Shanghai Securities News)

Pertanyaan Terkait

QWhat is the target IPO valuation and funding amount for SpaceX, and how does it compare to historical records?

ASpaceX is targeting a $75 billion IPO with a valuation of $1.75 trillion. This would be 2.5 times larger than the previous record holder, Saudi Aramco's $29.4 billion IPO, and exceeds the total of all U.S. IPOs in 2024 and 2025 combined.

QWhat is the approach of Starcloud in the AI infrastructure race and what is it unique?

AStarcloud is taking a unique approach by launching data centers into space. It completed a $170 million Series A, becoming Y Combinator's fastest unicorn, and used seed funding to launch a satellite carrying an H100 GPU. Its goal is a network of 88,000 satellites to create an orbital data center network.

QWhat role does Palantir play for the IRS, and what concerns have been raised about its work?

APalantir is working with the IRS to improve a tool called SNAP, which selects 'highest-value' targets for audits, tax collection, and criminal investigations. It is also building a 'mega API' to act as a central data interface. Democratic senators have raised concerns, calling it a 'surveillance nightmare'.

QAccording to the article, what is the current state of the conflict with Iran and its economic impact?

AThe conflict has entered its 31st day. Former President Trump has publicly advocated for invading Iran to 'seize oil,' and U.S. military units are deploying. Oil prices have surged over 50% to $116 per barrel, and the flow of oil through the Strait of Hormuz has dropped significantly from 20 million barrels per day, forcing Gulf countries to cut production.

QWhich Chinese AI company experienced a major outage, and what was the impact?

ADeepSeek experienced its longest outage since its launch, lasting 7 hours and affecting 355 million users. The cause of the outage was not publicly disclosed.

Bacaan Terkait

Model Generatif Bisa Dilatih End-to-End? Intinya Hanya Sebuah For Loop

Model generatif seperti model difusi atau autoregresif yang dominan saat ini biasanya tidak melatih seluruh proses generasi secara end-to-end. Mereka dilatih hanya untuk memprediksi satu "langkah kecil" dalam proses, namun saat inferensi, langkah ini harus diulang ratusan atau ribuan kali. Ketidaksesuaian ini menyebabkan masalah "exposure bias," di mana kesalahan menumpuk seiring waktu. Makalah terbaru dari UIUC dan Harvard memperkenalkan paradigma baru bernama **Explorative Modeling (XM)**. Inti dari metode ini sangat sederhana: alih-alih menghasilkan satu sampel, model menghasilkan **K kandidat** dalam setiap langkah pelatihan. Kemudian, hanya kandidat yang paling dekat dengan data target yang dipilih untuk menghitung gradien dan memperbarui model. Mengapa ini berhasil? Dalam tugas generatif, satu input bisa memiliki banyak output yang valid (mode). Fungsi kerugian rekonstruksi tradisional cenderung menyebabkan "mode blurring," di mana model hanya mempelajari rata-rata dari semua mode, menghasilkan output yang tidak realistis. Dengan menghasilkan beberapa kandidat, model dapat "mengeksplorasi" dan menangkap beberapa mode yang berbeda sekaligus, sehingga menghindari rata-rata yang kabur. Penelitian ini menunjukkan bahwa **"eksplorasi" (K) bertindak sebagai sumbu penskalaan ketiga** yang efektif, selain parameter dan data. Eksperimen pada gambar, video, dan bahasa menunjukkan peningkatan kinerja yang monoton dengan peningkatan K, dengan keuntungan yang lebih besar pada skala yang lebih besar. XM dapat meningkatkan efisiensi FLOP hingga 4.1x dan efisiensi sampel hingga 6.2x. Yang lebih menarik, ketika diterapkan secara ekstrem, XM memungkinkan **pelatihan generatif yang benar-benar end-to-end**. Dalam tugas kontrol robot, "Explorative Policy" berbasis XM mencapai kinerja yang setara dengan "Diffusion Policy" yang membutuhkan 100 langkah inferensi, sementara hanya membutuhkan satu langkah inferensi maju. Ini membuktikan bahwa kompleksitas untuk menangani banyak mode dapat dipindahkan dari inferensi ke pelatihan. Meskipun ide "best-of-K" bukan hal baru, kontribusi utama makalah ini adalah penjelasan teoritis bahwa mekanisme ini secara langsung meningkatkan **ekspresivitas generatif** model. Dengan demikian, ini menawarkan cara baru yang ampuh untuk mengatasi tantangan mendasar dalam pemodelan generatif.

marsbit6m yang lalu

Model Generatif Bisa Dilatih End-to-End? Intinya Hanya Sebuah For Loop

marsbit6m yang lalu

OpenAI Tidak Lagi Mengandalkan Model Termahal untuk Menghasilkan Uang

OpenAI tidak lagi mengandalkan model termahal untuk menghasilkan uang. Pada 30 Juli, perusahaan mengumumkan penyesuaian harga, menurunkan harga model GPT-5.6 Luna sebesar 80% dan Terra sebesar 20%. Yang lebih penting dari angka penurunan harga adalah pesan intinya: untuk banyak tugas, model terkuat (seperti Sol) tidak selalu diperlukan. OpenAI secara eksplisit merekomendasikan strategi di mana model mahal (Sol) menangani perencanaan dan analisis kompleks, sementara model yang lebih terjangkau (Luna) mengeksekusi tugas, menulis kode, dan menjalankan pengujian. Pergeseran ini mencerminkan perubahan industri AI dari hanya mengejar kecerdasan tertinggi menuju optimisasi biaya dan efisiensi untuk penggunaan skala besar. Anthropic juga mengikuti pola serupa dengan meluncurkan Claude Opus 5 dengan harga setengah dari model andalannya, Fable 5. Kedua raksasa Silicon Valley ini menunjukkan bahwa model flagship kini berfungsi lebih untuk nilai merek dan pembuktian teknologi, sementara model "kendaraan massal" yang lebih murah (seperti Luna, Terra, Opus) yang akan mendorong adopsi komersial luas dan menghasilkan pendapatan utama. OpenAI mengungkapkan bahwa penurunan biaya ini sebagian didorong oleh model AI (Sol) yang mengoptimalkan kode produksi dan alur kerjanya sendiri, menciptakan siklus peningkatan efisiensi yang dapat mempercepat penurunan biaya di masa depan. Intinya, kompetisi AI bergeser dari "siapa yang paling pintar" menjadi "mana yang paling bernilai". Perusahaan tidak lagi membeli satu model tunggal, tetapi merancang sistem dengan beberapa model yang cocok untuk tugas yang berbeda—mirip dengan arsitektur komputasi awan. Tujuan akhir OpenAI adalah membangun ekosistem yang mengunci alur kerja pengembang melalui volume panggilan API yang sangat besar dan biaya yang sangat rendah, sehingga menciptakan daya tarik yang kuat. Ketika AI menjadi murah dan tersebar luas seperti listrik, tertanam dalam setiap proses tanpa diperdebatkan, era sesungguhnya baru dimulai.

marsbit1j yang lalu

OpenAI Tidak Lagi Mengandalkan Model Termahal untuk Menghasilkan Uang

marsbit1j yang lalu

Trading

Spot
活动图片