Massive deficit, Sora2 service termination

OpenAI's AI video generation service "Sora 2" announced its shutdown on March 24, 2026. Against inference costs of approximately $15 million per day, cumulative in-app purchase revenue totaled a mere $2.1 million. Sora's creator Bill Peebles himself acknowledged that "the economics are completely unsustainable." For OpenAI as a whole, projected losses for 2026 stand at $14 billion, with HSBC warning of a "$207 billion funding shortfall." Amid this crisis, the fortunes of three key players are diverging sharply. Microsoft, having invested over $13 billion in OpenAI, is quietly distancing itself, spending $500 million annually on Anthropic to integrate Claude into its own products — a "diversified investment" strategy. It is also considering legal action after OpenAI allegedly breached its Azure contract by signing a deal worth over $50 billion with Amazon AWS — a cold-blooded strategy to protect its 27% stake (worth approximately $135 billion) acquired through its $13 billion investment, while reducing dependence on OpenAI. By contrast, SoftBank Group has poured a cumulative $61 billion into OpenAI and plans to invest $500 billion through the Stargate project, yet was scrambling to manage $22.5 billion in cash flow as recently as December 2025. Its all-in bet on a successful OpenAI IPO evokes memories of WeWork during the Vision Fund era. As for OpenAI itself, it faces a 4x cost disadvantage against Google TPUs due to its Nvidia GPU-dependent infrastructure, has been overtaken by Anthropic in enterprise market share (27% vs. 40%), and is on the verge of a legal dispute with its largest backer, Microsoft. The shutdown of Sora 2 is not merely a product withdrawal — it is a symbol of a tectonic shift underway in the AI industry.

Summary

Microsoft is quietly distancing itself from OpenAI — into which it poured over $13 billion — while spending $500 million annually on Anthropic to integrate Claude into its own products, pursuing a "diversified investment" strategy. It is also considering legal action after OpenAI betrayed its Azure contract by signing a deal worth over $50 billion with Amazon AWS — a cold-blooded strategy to protect the 27% stake (worth approximately $135 billion) it obtained through its $13 billion investment, while reducing its dependence on OpenAI.

SoftBank Group, by contrast, has committed a cumulative $61 billion to OpenAI and is planning $500 billion through the Stargate project, yet was scrambling to manage $22.5 billion in cash flow as recently as December 2025. Its all-in bet on a successful OpenAI IPO evokes memories of WeWork during the Vision Fund era.

OpenAI itself operates infrastructure dependent on Nvidia GPUs that carry a 4x cost disadvantage relative to Google TPUs, has seen its enterprise market share reversed by Anthropic — 27% to 40% — and is on the verge of legal conflict with its largest backer, Microsoft. The shutdown of Sora 2 is not merely a product withdrawal; it is a symbol of a tectonic shift reshaping the AI industry.

Text

Sora 2 — Outpouring of Dissatisfaction Over Quality and Speed, Service Ends After Just 6 Months

OpenAI's Sora is an AI model that generates video from text. It was first unveiled as a research demo in February 2024 and made generally available on December 9 of the same year. "Sora 2," launched on September 30, 2025, featured improved physical accuracy, video generation of 15–25 seconds, synchronized audio generation, and a "Cameo" function that inserts selfie videos into generated scenes, along with a standalone iOS app and API.

However, user complaints were constant from the beginning. Wait times for video generation were long, and timeouts occurred frequently during peak load hours. The quality of generated videos fell short of expectations set by demo footage, with ongoing criticism of unnatural depictions of human hands and physical movement. The app itself ran sluggishly, and complaints about usability flooded social media and review sites. Download numbers peaked at approximately 3.3 million in November 2025, then plummeted to roughly 1.1 million by February 2026.

On March 24, 2026, OpenAI announced the end of Sora 2. All video generation features across the iOS app, API, sora.com, and ChatGPT will be discontinued.

According to analysis by Deepak Mathivanan of Cantor Fitzgerald, Sora 2's inference costs reached approximately $15 million per day (roughly ¥2.25 billion), or about $5.4 billion (roughly ¥810 billion) annually. Meanwhile, cumulative in-app purchase revenue totaled only about $2.1 million (roughly ¥315 million), with monthly revenue below $500,000 (roughly ¥75 million). Bill Peebles, the creator of Sora himself, acknowledged in October 2025 that "the economics are completely unsustainable." Forrester's Thomas Husson called it a "resource black hole," and KeyBanc's Justin Patterson analyzed that "downloads and engagement have decayed." Nitish Pahwa of Slate wrote that "Sam Altman is finally acknowledging what no one wanted to admit."

OpenAI's Decline — $14 Billion (roughly ¥2.1 Trillion) in Losses and a $207 Billion (roughly ¥31 Trillion) Funding Gap

The end of Sora 2 is just one facet of OpenAI's broader financial crisis.

In the first half of 2025 alone, the company recorded $13.5 billion (roughly ¥2.025 trillion) in losses against $4.3 billion (roughly ¥645 billion) in revenue, with $2.5 billion (roughly ¥375 billion) in cash burn. Against projected 2026 revenue of approximately $13 billion (roughly ¥1.95 trillion), projected losses stand at $14 billion (roughly ¥2.1 trillion)—losses exceeding revenue. Cash burn in 2027 is projected at $57 billion (roughly ¥8.55 trillion), operating losses in 2028 at $74 billion (roughly ¥11.1 trillion), and cumulative cash burn through 2029 is expected to reach $115 billion (roughly ¥17.25 trillion).

An analyst team at HSBC concluded that "OpenAI is unlikely to become profitable by 2030 and faces a $207 billion (roughly ¥31.05 trillion) funding gap." Inference costs quadrupled in 2025, and gross margins deteriorated from 40% to 33%.

Fidji Simo, CEO of OpenAI's applications division, declared at an all-hands meeting: "We can't afford to get distracted by side quests and miss this." VC and analyst Om Malik, writing under the headline "OpenAI Has New Focus (on the IPO)," pointed out that Sora's termination is a stepping stone toward creating a "cleaner revenue story" ahead of the planned IPO (scheduled for Q4 2026).

Microsoft Stays — Protecting $13 Billion+ (roughly ¥1.95 Trillion+) in Investment While Reducing OpenAI Dependence

Amid this turmoil, Microsoft is executing an exceptionally cold-eyed strategy.

Microsoft's cumulative OpenAI investment exceeds $13 billion (roughly ¥1.95 trillion). With OpenAI's conversion to a for-profit company in October 2025, Microsoft secured a 27% stake (worth approximately $135 billion, or roughly ¥20.25 trillion) and an Azure contract worth $250 billion (roughly ¥37.5 trillion). At the same time, however, the exclusive cloud rights it once held were lost.

Microsoft is already spending $500 million per year (roughly ¥75 billion) on Anthropic and has integrated Claude models into products such as Microsoft 365 Copilot. In September 2025, TechCrunch reported that "Microsoft is buying AI from Anthropic to reduce its OpenAI dependence," and 247 Wall Street analyzed whether "Microsoft's $500 million AI pivot to Anthropic is an admission of failure."

Microsoft's "best model for the job" approach goes beyond a mere technical judgment. While OpenAI's enterprise market share has fallen from 50% to 27%, Anthropic has risen to 40%, and among first-time AI adopters, Anthropic wins roughly 70% of head-to-head comparisons. Microsoft is picking its winner.

From Azure to AWS — OpenAI's $50 Billion+ (roughly ¥7.5 Trillion+) "Betrayal" and Microsoft's Fury

What decisively soured the Microsoft–OpenAI relationship was OpenAI's move to Amazon AWS.

OpenAI struck a $38 billion (roughly ¥5.7 trillion) cloud deal with AWS at the end of 2025, later expanding it to $100 billion (roughly ¥15 trillion) over eight years. AWS became the exclusive third-party cloud provider for OpenAI's enterprise platform "OpenAI Frontier" (launched February 2026). For Microsoft, this was tantamount to trampling on the spirit of the $250 billion Azure contract.

According to WinBuzzer reporting, Microsoft is considering legal action against this $50 billion+ (roughly ¥7.5 trillion+) deal. The legal argument centers on the distinction between "stateless API calls" under the Azure contract and Amazon's "stateful runtime environment." OpenAI itself has listed "dependence on Microsoft" as a risk factor in pre-IPO investor documents—an extraordinary situation in which the company labels the partner who invested in and supported it as a "risk."

Behind this dynamic lies the issue of infrastructure costs. Google's TPU v6e offers four times better cost-per-performance than Nvidia's H100. Through its own chips, Google procures AI compute at roughly one-fifth the cost of Nvidia GPU purchasers. While Nvidia's GPU manufacturing cost runs $3,000–$5,000 (roughly ¥450,000–¥750,000), hyperscalers pay $20,000–$35,000 (roughly ¥3 million–¥5.25 million) or more per unit—the "Nvidia tax." TPUs also consume 60–65% less power.

By migrating to TPUs, Midjourney cut its monthly inference costs from $2.1 million (roughly ¥315 million) to below $700,000 (roughly ¥105 million). Anthropic has signed the largest TPU contract in history with Google, securing 1 million Trillium TPUs by 2027. OpenAI, locked into Nvidia GPU-dependent Azure and losing cost competitiveness, sought an outlet in AWS—the result being legal conflict with Microsoft.

Microsoft's position is ironic but rational. The roughly ¥20 trillion value of its 27% stake—acquired through $13 billion+ in investment—is worth preserving. Maintaining that stake while hedging technology risk through diversified investment in Anthropic, and legally contesting OpenAI's AWS deal: the shift from "shared destiny" to "one holding in a portfolio" is, in the end, the most risk-managed strategy on the table.

SoftBank's Descent — A $61 Billion (roughly ¥9.15 Trillion) All-In Bet

By contrast, SoftBank Group's position is precarious.

SoftBank's cumulative commitment to OpenAI has reached $61 billion (roughly ¥9.15 trillion): $500 million (roughly ¥75 billion) in October 2024, $30 billion (roughly ¥4.5 trillion) in March 2025, and $30 billion (roughly ¥4.5 trillion) in February 2026. It is also a core partner in the "Stargate" project—a joint venture among OpenAI, SoftBank, Oracle, and MGX targeting up to $500 billion (roughly ¥75 trillion) in U.S. AI infrastructure investment through 2029.

However, according to SiliconANGLE reporting, in December 2025 SoftBank was scrambling to secure $22.5 billion (roughly ¥3.375 trillion) in funding, reportedly exploring borrowing $11.5 billion (roughly ¥1.725 trillion) using Arm shares as collateral and tapping its $11 billion (roughly ¥1.65 trillion) stake in T-Mobile. Equity funding accounts for only 10% of the total, with the remainder dependent on debt. Bloomberg reported in May 2025 that the Stargate project was slowing due to tariff concerns.

While Microsoft diversifies toward Anthropic with a "best model for the job" approach, SoftBank is betting everything on OpenAI's success. The outcome of OpenAI's IPO (scheduled for Q4 2026) holds the key to SoftBank's investment recovery. Yet that same OpenAI is on track for $14 billion (roughly ¥2.1 trillion) in losses in 2026, has seen its enterprise market share shrink to 27%, and is on the brink of legal conflict with its largest partner, Microsoft. VC and analyst Tomasz Tunguz frames OpenAI's IPO as part of a "$3 trillion stress test," questioning whether public markets can absorb a company of this scale.

The memory of concentrated WeWork bets during the Vision Fund era—resulting in roughly ¥450 billion in valuation write-downs—remains vivid in the investor community. The difference between OpenAI and WeWork is that OpenAI's technology is real—but "having real technology" and "being able to monetize it" are separate questions.

The Full Picture of Investors — $168 Billion (roughly ¥25.2 Trillion) in Investment and Fraying Relationships

OpenAI has raised a cumulative total of approximately $168 billion (roughly ¥25.2 trillion) across 11 rounds from 68 investors.

In the February 2026 round of $110 billion (roughly ¥16.5 trillion, at a valuation of $730 billion or roughly ¥109.5 trillion), Amazon contributed $50 billion (roughly ¥7.5 trillion, with a first tranche of $15 billion or roughly ¥2.25 trillion), Nvidia contributed $30 billion (roughly ¥4.5 trillion), and SoftBank contributed $30 billion (roughly ¥4.5 trillion, with a first tranche of $10 billion or roughly ¥1.5 trillion).

In the October 2024 round of $6.6 billion (roughly ¥990 billion, at a valuation of $157 billion or roughly ¥23.55 trillion), Microsoft invested approximately $1 billion (roughly ¥150 billion), Thrive Capital approximately $1.3 billion (roughly ¥195 billion), Khosla Ventures over $500 million (roughly over ¥75 billion), and Nvidia $100 million (roughly ¥15 billion).

Impact on the Industry

The end of Sora 2 has thrown three contrasting fortunes into sharp relief.

A sinking OpenAI — $14 billion in losses in 2026, a $207 billion funding shortfall, enterprise share shrinking to 27%, legal exposure from dual contracts with Microsoft and AWS, and a structural cost disadvantage against Google TPUs. Shuttering Sora 2 is cleanup ahead of an IPO — but the fact that so much needed cleaning up is the real problem.

A sinking SoftBank G — a $61 billion all-in bet on OpenAI, $500 billion Stargate, and $22.5 billion in cash-flow strain. If OpenAI's IPO fails, it will be a replay of the Vision Fund era. The greatest risk is having no option to diversify.

A surviving Microsoft — over $13 billion invested to lock in a 27% stake (worth roughly $200 billion), while hedging with $500 million a year into Anthropic and fighting OpenAI's AWS contract on legal grounds. The cold-blooded pivot from "shared destiny" to "one asset in the portfolio" has turned out to be the most risk-managed strategy of all. Microsoft's real strength is not that it invests in AI companies — it's that no matter how many AI companies fail, its own cloud and enterprise business will survive.

Sources: Axios, "OpenAI will shutter Sora video app" (March 24, 2026); CNBC, "OpenAI shutters Sora as company reels in costs" (March 24, 2026); Bloomberg, "OpenAI discontinues Sora" (March 24, 2026); Variety, "OpenAI shutting down Sora; Disney drops $1B plan" (March 2026); Fortune, "OpenAI cash burn rate annual losses through 2028" (November 2025); Yahoo Finance, "OpenAI's own forecast predicts $14B loss in 2026"; HSBC analysts OpenAI profitability forecast; Deepak Mathivanan (Cantor Fitzgerald) Sora cost analysis; Thomas Husson (Forrester) "resource black hole"; Justin Patterson (KeyBanc Capital Markets) engagement analysis; Bill Peebles (Head of Sora) "economics unsustainable" (October 2025); Nitish Pahwa (Slate), "Sam Altman Is Finally Admitting Something" (March 2026); Om Malik, "OpenAI Has New Focus (on the IPO)" (March 2026); Tomasz Tunguz IPO stress test analysis; 247 Wall Street, "Microsoft's $500M AI pivot to Anthropic" (January 2026); TechCrunch, "Microsoft to lessen reliance on OpenAI" (September 2025); WinBuzzer, "Microsoft weighs suing OpenAI over Amazon cloud deal" (March 2026); SiliconANGLE, "SoftBank scrambles to find $22.5B" (December 2025); Bloomberg, "SoftBank Stargate snags on tariff fears" (May 2025); Nasdaq, "Google TPU advantage vs OpenAI's Nvidia tax"; Cybernews, "Why did OpenAI's Sora crash and burn?"