OpenAI's $500B Stargate Week: Everything Hitting at Once

Summary: In the final days of January 2025, OpenAI was simultaneously announcing the largest AI infrastructure pledge in history, actively renegotiating its Microsoft partnership, pushing to convert its legal structure from nonprofit-controlled to for-profit, defending against escalating European privacy enforcement, and watching a Chinese lab threaten its pricing power — all in the same week. This article unpacks each of those threads with precise figures, dates, and implications for developers, enterprises, and policymakers, and clearly separates what was confirmed that week from what was still under negotiation and only finalized later.

On January 29, 2025, science-fiction author John Scalzi posted just three words to Bluesky — "OpenAI right now" — and the internet understood immediately. It was the kind of shorthand that lands only when a company is simultaneously everywhere and under fire from every direction. That single week, OpenAI right now meant a $500 billion infrastructure pledge, a rattled stock market, a restructuring battle, a brewing privacy fight, and a partnership with Microsoft that was being rewritten in real time. Here is the full picture.

The Stargate Announcement and Why It Shook Everything

On January 21, 2025 — eight days before Scalzi's post — President Donald Trump stood in the White House to announce The Stargate Project, a joint venture among OpenAI, Oracle, SoftBank, and MGX targeting $500 billion in AI infrastructure investment over four years. The initial commitment was reported at $100 billion, earmarked for data centers and compute capacity across the United States, with OpenAI serving as the operational lead and SoftBank leading the financing. SoftBank's Masayoshi Son and Oracle's Larry Ellison appeared alongside Trump and OpenAI CEO Sam Altman, lending the announcement a theatrical scale that matched its financial one.

The scale was staggering by any historical measure of private technology investment. But within days, it collided head-on with news from China: DeepSeek's R1 model, released publicly in late January 2025, demonstrated reasoning capabilities competitive with OpenAI's best models at a fraction of the reported training cost. The juxtaposition — OpenAI pledging half a trillion dollars to build more compute just as a Chinese lab argued you might need far less — helped send Nvidia's stock into its largest single-day market-capitalization loss in U.S. equity history on January 27, 2025 (roughly $590 billion erased in one session), and sparked furious debate about whether America's compute-maximalist AI strategy was fundamentally mispriced.

For developers and technical readers, the DeepSeek moment raised a pointed question: if frontier reasoning can be distilled and replicated cheaply, does the best OpenAI right now moat rest on compute scale, proprietary data, safety research, distribution reach, or simply brand? There was no clean answer in January 2025 — which is precisely why Scalzi's three-word post needed no explanation.

The Microsoft Partnership Rewrite: Right of First Refusal, Equity, and Revenue Sharing

Running beneath the Stargate headline was a quieter but arguably more consequential story: OpenAI and Microsoft were fundamentally renegotiating the terms of their relationship. Importantly, in the January 2025 window itself these were negotiations and proposals, not settled terms — the definitive restructured agreement (with the specific equity splits and the surrender of Microsoft's cloud right of first refusal) was only completed later, on October 28, 2025. For anyone evaluating the Microsoft OpenAI right of first refusal arrangement, that later agreement marks the clear before-and-after moment, and the January 2025 negotiations were where its outline first became visible.

Microsoft had invested a reported total of more than $13 billion into OpenAI across several tranches, securing deep Azure integration rights and — critically — a right of first refusal over OpenAI's future cloud infrastructure purchases. That clause gave Microsoft structural leverage: whenever OpenAI needed to expand its compute footprint, Microsoft got the first opportunity to win that contract. The Stargate structure — which brought Oracle and SoftBank in as compute financiers and operators — could only proceed if that exclusivity was loosened. In the definitive agreement announced later in 2025, Microsoft agreed to give up that right of first refusal on new compute, opening the door to Oracle and other providers joining OpenAI's compute stack.

Under that finalized arrangement, Microsoft held an approximately 27% equity stake in the restructured for-profit entity (OpenAI Group PBC), which Microsoft valued at roughly $135 billion at the time it was announced. OpenAI, for its part, continued to designate Azure as a primary cloud provider under a large ongoing purchase commitment (Microsoft cited an incremental Azure commitment on the order of $250 billion), and the two companies maintained a revenue-sharing arrangement, with an independent expert panel tasked with verifying any future declaration that artificial general intelligence had been achieved — a trigger that would subject several contractual terms to change.

It is worth stressing the timeline: as of Scalzi's late-January 2025 post, the exact equity percentages, valuation, ROFR surrender, and Azure commitment value had NOT been publicly confirmed as binding. The figures circulating that week were reported estimates and negotiating positions. This article distinguishes those from the terms that were later finalized.

Term	Pre-Restructuring (capped-profit era)	Post-Restructuring (finalized Oct 2025)
Microsoft economic interest	Right to up to ~49% of the capped-profit LLC's profits (not an equity stake)	~27% equity stake in OpenAI Group PBC
Right of first refusal (cloud)	Microsoft held it	Microsoft gave it up on new compute
Revenue sharing	Revenue share to Microsoft (reported ~20%)	Revenue share continues; terms tied to AGI verification
Azure commitment	Exclusive primary provider	Continued primary provider; Oracle/SoftBank join Stargate stack
Third-party cloud flexibility	Restricted by ROFR	Loosened — Oracle and others now eligible

The restructuring of these terms matters enormously for developers building on OpenAI's APIs. A more diversified infrastructure base — spanning Azure, Oracle Cloud, and potentially additional providers — implies greater redundancy and capacity headroom, but also raises real questions about latency consistency and feature parity across hosting environments. It is also a signal that OpenAI increasingly sees itself as a product company that happens to need cloud infrastructure, rather than a Microsoft subsidiary that happens to build models. That shift in posture has direct implications for every enterprise software team evaluating long-term AI vendor lock-in. You may also want to read about how intellectual property risk shapes AI vendor relationships as you assess OpenAI's evolving independence.

The Corporate Restructuring: From Nonprofit-Controlled to Public Benefit Corporation

To understand the legal and governance turbulence swirling around OpenAI in January 2025, you need to understand what the company was in the middle of changing. OpenAI was founded in 2015 as a nonprofit. It later created a capped-profit subsidiary to attract investment while nominally keeping the mission intact — a structure that limited investor returns to a multiple of their principal, with surplus value theoretically reverting to the nonprofit parent. By late December 2024, the company had publicly outlined a plan to convert that subsidiary into a Delaware public benefit corporation (PBC), a move that would remove the profit cap and reduce the nonprofit's direct operational control over the for-profit business.

As eventually finalized in October 2025, OpenAI Group PBC would operate with the OpenAI Foundation (the surviving nonprofit) holding roughly a quarter of the company — reported at about 26% — Microsoft at approximately 27%, and current and former employees plus other investors holding the remainder. As of January 2025, however, these were proposed figures under active negotiation and regulatory review; nothing about the equity split or the Foundation's stake was settled. The nonprofit's stake represented, in effect, its compensation for releasing governance control — a valuation exercise that drew pointed scrutiny from nonprofit law scholars, former employees, and state attorneys general.

Critics argued that the nonprofit, created with charitable assets and carrying tax-exempt status, was surrendering mission-driven governance in exchange for a financial stake in a company it would no longer fully control. Some former employees and nonprofit watchdog groups went public with their concerns. OpenAI, according to reporting at the time, issued subpoenas to individuals and organizations it believed were coordinating opposition — a move that deepened the perception of a company more concerned with suppressing external opposition than engaging it.

Regulatory approval was a live question as of January 2025. The offices of the Delaware and California attorneys general — the latter especially significant given that OpenAI operates primarily out of San Francisco — were both reviewing the conversion, and neither had signed off. Any characterization of those approvals as settled at that point would be premature; both ultimately reached memoranda of understanding with OpenAI later in 2025 before the restructuring closed.

Separately, Elon Musk, who had co-founded OpenAI and departed its board in 2018, had brought litigation (originally filed in 2024 and subsequently refiled and amended) seeking to block or reverse the nonprofit-to-PBC conversion, alleging that the restructuring betrayed the charitable mission under which OpenAI originally solicited donations and talent. The litigation was ongoing as of early 2025.

Why it matters for developers: A PBC structure means OpenAI's directors must balance shareholder interests against its stated public benefit mission. In practice, this creates a governance buffer against pure profit-maximization — but a substantially weaker one than direct nonprofit board control over operations. Developers and enterprises building long-term products on OpenAI's infrastructure should understand that the entity they are contracting with is, increasingly, a conventional technology corporation operating under a public benefit charter. The mission language remains; the structural enforcement mechanism has been materially weakened.

OpenAI Right to Be Forgotten: The Privacy Challenge Gets Sharper

Alongside the corporate drama, a quieter but legally significant issue was intensifying in early 2025: the OpenAI right to be forgotten challenge. Under the European Union's General Data Protection Regulation (GDPR), individuals hold the right to request erasure of personal data held about them. For a large language model trained on scraped web data, fulfilling that right is technically non-trivial — you cannot simply delete a database row when personal information has been absorbed into billions of model weights as statistical patterns rather than stored records.

European regulators had been sharpening their focus on exactly this question for over two years. Italy's Garante — the national data protection authority — forced a temporary ChatGPT service suspension in March 2023, restoring access only after OpenAI implemented specific disclosure and opt-out mechanisms; the Garante later, in December 2024, fined OpenAI €15 million over its data practices. By early 2025, the Irish Data Protection Commission (acting as OpenAI's lead EU supervisory authority under GDPR's one-stop-shop mechanism, following OpenAI's establishment of a European base in Ireland) and other national regulators were applying broader enforcement frameworks with increased consistency and coordination.

The technical tension is stark: an individual's name, medical history, or financial details, if present in training data, become entangled with the model's statistical structure in ways that cannot be surgically excised without either full retraining or targeted fine-tuning to suppress specific memorized sequences — both expensive and imperfect processes. Techniques such as machine unlearning were an active area of research, but no production-grade solution for large-scale GDPR erasure requests had been demonstrated at the scale ChatGPT required.

For OpenAI, this creates a compounding liability as its user base expands. ChatGPT was reported by multiple outlets to have surpassed roughly 20 million paid subscribers around late 2024, a figure that grew further into 2025, alongside a substantial base of enterprise API customers. Distinguishing between data processed to improve models versus data processed only during inference is a compliance boundary that EU regulators were demanding OpenAI make far more legible. Enterprises relying on OpenAI's API for workflows involving personal data should treat this regulatory trajectory as a material risk factor — a point developed further in our coverage of why uploading medical records to any LLM is a privacy disaster.

What Is the Best OpenAI Model Right Now? The Developer's Guide to the January 2025 Lineup

Setting aside governance and geopolitics, developers asking practically — what is the best OpenAI right now for my use case? — were navigating a model lineup that had expanded significantly in late 2024 and early 2025, with meaningfully different performance-cost trade-offs across tiers.

The o-Series Reasoning Models: o1 and o3

OpenAI's o1 model, released in September 2024 and broadly available by late 2024, represented the company's first public deployment of a "chain-of-thought at inference time" architecture — a model that generates extended internal reasoning traces before producing a final output, rather than predicting answers in a single autoregressive pass. This approach produced measurably higher accuracy on complex mathematical, coding, and multi-step logical tasks at the cost of higher latency and per-token expense.

o3, announced in December 2024, extended this reasoning architecture with significantly deeper compute budgets during inference. The benchmark result that electrified — and genuinely unsettled — the AI research community was o3's performance on ARC-AGI (the Abstraction and Reasoning Corpus for Artificial General Intelligence), a benchmark specifically designed to resist pattern memorization and test novel reasoning. On the semi-private evaluation set, o3 scored approximately 87.5% in its high-compute configuration (about 75.7% in the lower-compute setting), compared with roughly 5% for GPT-4o — a jump so large it prompted ARC-AGI co-creator François Chollet to publicly call it a genuine breakthrough on the benchmark. o3-mini, a smaller and more cost-accessible variant, launched on January 31, 2025; full o3 access for developers followed in subsequent months.

GPT-4o and the Production Workhorse Tier

GPT-4o remained the dominant choice for most production applications as of January 2025: meaningfully faster and cheaper than o1 for tasks that do not require extended reasoning chains, with strong multimodal capabilities spanning text, image input, and audio. The practical decision rule for most teams was simple: use GPT-4o for latency-sensitive and cost-constrained applications; reach for o1 or o3 where accuracy on complex reasoning problems justifies the inference cost premium.

The DeepSeek R1 release complicated this calculus by demonstrating competitive reasoning performance via an open-weights model (released under an MIT license) that organizations could self-host at near-zero marginal inference cost. This put immediate downward pricing pressure on OpenAI's o-series tier and accelerated developer experimentation with open-weight alternatives — worth comparing with how MiniCPM5-1B achieves 128K-token reasoning on a single GPU as a concrete example of how capable smaller, self-hostable models had become.

Model	Best For	Relative Cost	Reasoning Depth	Context Window
GPT-4o	Production apps, multimodal, low latency	Moderate	Standard	128K tokens
o1	Complex reasoning, code, math	High	Extended (chain-of-thought)	200K tokens
o3 / o3-mini	Frontier reasoning, research, hard benchmarks	Very High (o3) / Moderate (o3-mini)	Deep (multi-step, extended compute)	200K tokens
GPT-4o mini	High-volume, cost-sensitive pipelines	Low	Basic	128K tokens

The Revenue Picture and What It Funds

Beneath all the noise, OpenAI's financial trajectory in early 2025 was objectively remarkable — and objectively incomplete as a business story. The company exited 2024 with annualized revenue reported at roughly $3.7 billion, driven primarily by ChatGPT subscriptions and API usage. Subsequent reporting indicated this figure would scale toward roughly $12 billion annualized during 2025. The April 2025 funding round — up to $40 billion at a reported $300 billion post-money valuation, led by SoftBank — would, if fully closed on the reported terms, rank as the largest single private technology fundraise on record. These figures are cited here as forward context; they were not yet public knowledge at the moment of Scalzi's January post.

That capital is being directed toward three buckets: the Stargate compute buildout, safety and alignment research (OpenAI's stated justification for remaining at the frontier rather than ceding it to actors with less stated commitment to safety), and the product surface area — ChatGPT, the API ecosystem, enterprise integrations — that generates the revenue. The company's core argument, implicit in every major announcement of this period, is that scale and safety are complements rather than trade-offs: that only a well-resourced OpenAI can afford the safety work that frontier AI requires. That claim remains genuinely contested in the AI research community, and the DeepSeek episode made it harder to sustain without qualification, since it suggested frontier performance might not require the resource scale OpenAI was assembling.

OpenAI remained loss-making through this period — infrastructure costs, talent compensation, and research spending significantly outpaced income. The path to profitability depended heavily on continued subscriber and enterprise growth, sustained API pricing power, and the compute cost reductions that hardware improvements and architectural efficiency gains were expected to deliver over the next several years.

The healthcare sector's engagement with AI tools was also accelerating through this period, even as regulatory and safety scrutiny intensified — a dynamic that connects directly to concerns raised by frontline healthcare workers about AI surveillance in clinical settings.

Frequently Asked Questions

What is OpenAI right now, in terms of company structure?

As of early 2025, OpenAI was in the process of converting from a capped-profit LLC controlled by a nonprofit parent into OpenAI Group PBC, a Delaware public benefit corporation. Under the plan later finalized in October 2025, the OpenAI Foundation (the surviving nonprofit) held roughly 26% of the new entity, Microsoft approximately 27%, and employees plus other investors the remainder. As of January 2025, these were proposals under active negotiation and regulatory review — not settled terms.

What is the best OpenAI model right now for developers?

For most production workloads in early 2025: GPT-4o for speed and cost efficiency with strong multimodal support; o1 for complex reasoning tasks where accuracy justifies the latency and cost premium; o3 (via o3-mini from January 31, 2025 onward) for the highest-difficulty reasoning benchmarks; and GPT-4o mini for high-throughput, cost-sensitive pipelines. The right choice depends heavily on task complexity, latency tolerance, and budget.

What does the Microsoft OpenAI right of first refusal mean?

The right of first refusal gave Microsoft the contractual right to be the first provider offered new OpenAI cloud infrastructure contracts. As part of the restructuring negotiations that began to surface in January 2025 and were finalized on October 28, 2025, Microsoft agreed to give up this right on new compute, allowing OpenAI to bring Oracle and other cloud providers into its infrastructure stack. Microsoft retained an equity stake (reported at ~27%) and Azure's position as a primary cloud provider.

What is the OpenAI right to be forgotten, and why is it hard to implement?

Under GDPR, individuals can request deletion of their personal data. For OpenAI, this is technically complex because personal information present in training data is not stored as discrete records — it is encoded as statistical patterns across billions of model weights. Fulfilling a deletion request requires either full model retraining or targeted fine-tuning to suppress memorized content, both of which are expensive, slow, and imperfect. EU regulators were actively pressing OpenAI on this issue in early 2025, following Italy's Garante's 2023 enforcement action and its December 2024 €15 million fine.

How did DeepSeek R1 affect OpenAI's position?

DeepSeek R1's late-January 2025 release demonstrated reasoning performance competitive with OpenAI's o1 at a reported training cost dramatically lower than frontier norms, using an open-weights (MIT-licensed) architecture that organizations could self-host. This challenged the premise of OpenAI's Stargate compute investment, pressured its o-series pricing, and accelerated enterprise evaluation of open-weight alternatives. It did not displace OpenAI's product distribution advantages, but it narrowed the purely technical moat on reasoning performance.

Key Takeaways

• Stargate is the defining bet: OpenAI committed to a $500 billion AI infrastructure joint venture with Oracle, SoftBank, and MGX on January 21, 2025 — days before DeepSeek R1 publicly questioned the compute-maximalist thesis underlying that bet.
• Microsoft's cloud right of first refusal was on the table: the negotiations visible in January 2025 pointed toward Microsoft giving up its right of first refusal on new compute, retaining roughly 27% equity and Azure's status as a primary provider — terms that were finalized on October 28, 2025, opening the door for Oracle and others in the Stargate stack.
• The nonprofit-to-PBC conversion fundamentally changes OpenAI's governance: under the plan finalized later in 2025, the OpenAI Foundation retains roughly 26%, but operational control shifts to a for-profit public benefit corporation. As of January 2025 this remained a proposal, subject to still-pending state attorney general reviews and contested in litigation brought by Elon Musk.
• The OpenAI right-to-be-forgotten challenge is unresolved and escalating: EU regulators are pressing OpenAI on GDPR erasure rights that are technically difficult to fulfill for a model trained on web-scale data, following Italy's 2023 enforcement action and its December 2024 fine.
• For the best OpenAI model right now: GPT-4o is the production standard; o1 and o3 (via o3-mini from January 31, 2025) form the frontier reasoning tier; GPT-4o mini serves high-volume, cost-sensitive pipelines. DeepSeek R1 has introduced real competitive pressure on reasoning-tier pricing specifically.
• Revenue is real and growing fast: approximately $3.7 billion annualized at end-2024 — but the company remains loss-making as infrastructure costs and talent spending significantly outpace current income.

What Comes Next

The months ahead for OpenAI will be shaped by three converging forces. First, whether the Stargate compute buildout can actually be executed at the claimed pace: real-estate permitting, power grid access, transformer procurement lead times, and advanced chip supply-chain constraints are practical limiters that no announcement resolves. Second, whether the PBC restructuring survives legal and regulatory challenge — Musk's lawsuit was actively amended in the months following January 2025, and state attorney general reviews in both California and Delaware remained live proceedings before both reached agreements enabling the October 2025 close. Third, whether the model capability lead that justifies OpenAI's premium pricing holds as open-weight models from DeepSeek, Meta, Mistral, and others continue to close the performance gap at lower cost.

John Scalzi's three words landed because they captured a company operating simultaneously at maximum ambition and maximum vulnerability — making the largest infrastructure bet in the industry's history in the same week a competitor argued the bet's core assumption might be wrong, pushing to convert its legal structure in the face of regulatory and legal opposition, and defending its training-data practices against regulators on two continents. That tension is not resolving anytime soon. Understanding it clearly is the prerequisite for any serious decision about building on, investing in, or competing with OpenAI.

---

Figures and dates in this article draw on contemporaneous public reporting and primary sources including OpenAI's and Microsoft's own announcements, the ARC Prize evaluation of o3, and court filings in the Musk litigation. Where figures were estimates or subject to ongoing negotiation as of late January 2025 — particularly the Microsoft equity percentage, the Azure commitment value, and the proposed restructuring equity splits — this is noted in the text, and terms that were only finalized later (notably the definitive Microsoft/OpenAI restructuring completed October 28, 2025) are labeled as such. Forward-looking figures (the April 2025 fundraise and 2025 revenue projections) are cited as subsequent context and were not publicly known as of January 29, 2025.