OpenAI’s audited tab for 2025 came in at roughly $34 billion, a figure that has turned a hotly anticipated listing into a referendum on whether AI economics can ever match AI ambition. With $19 billion tagged for research and development and nearly $6 billion for sales and marketing, the rest landing on infrastructure and headcount, the company is leaning into scale ahead of a potential IPO that some reports say could aim for a valuation near $1 trillion. The bet lands as a coalition of 42 state attorneys general opens a broad probe into OpenAI’s practices, and as investors handicap spillover effects for Microsoft (MSFT) and Nvidia (NVDA).
The headline number is the point. In a year when investors rewarded companies that proved operating leverage, OpenAI pushed in the opposite direction: spend now to dominate later. The audited breakdown reinforces where the arms race is hottest. R&D towers over the P&L at $19 billion, a marker of the cost to train and refine frontier models and to ship a drumbeat of upgrades that keep usage high. Nearly $6 billion on sales and marketing underscores a land-grab in enterprise, where procurement cycles are long, security reviews are punishing, and incumbents such as Microsoft and Google maintain distribution advantages. What the audited figures do not solve is the central IPO puzzle: how much revenue sits on the other side of that spend, and at what gross margin once cloud bills and inference costs clear. Without a clear unit economics story, $34 billion reads as both commitment and liability.
OpenAI does not own a hyperscale cloud. It rents one. Microsoft’s multiyear partnership gives OpenAI access to vast compute on Azure, but it also embeds the company’s cost structure inside someone else’s capital plan. That is great for speed to market and catastrophic if GPU prices, power costs, or contract terms move against you. Training is lumpy capex-by-proxy. Inference is recurring COGS. The two together define the “Nvidia tax” that every model builder pays until custom silicon or radical efficiency gains arrive. Pricing power in AI services remains unproven beyond premium enterprise tiers. If competition forces vendors toward usage-based discounts or bundles, margins can vanish quickly. For a public market audience, the critical tables in an S-1 will be pre-purchased compute obligations, average cost per token for flagship models, and how fast those figures fall with each generation. If the slope is shallow, valuation math tightens.
The multistate investigation, led by New York Attorney General Letitia James, hits on the fault lines that matter for any consumer-facing AI platform: ads and claims, data handling for minors and sensitive health information, engagement mechanics that could spur overuse, safety protocols, and model behavior. It arrives days after reports of a confidential IPO filing, tightening the disclosure window. OpenAI said it takes the concerns “seriously” and signaled cooperation. That will not stop regulators from demanding documentation, audits, and possibly product changes. For investors, the issue is not just fines. It is product friction and monetization limits. If ad copy, onboarding flows, or data retention policies need to shift, the growth engine supporting those sales and marketing dollars turns. And in a market newly intolerant of governance surprises, a still-unresolved probe will widen the valuation spread between bull and bear cases.
Microsoft is OpenAI’s essential counterparty and chief distribution channel. Azure runs the workloads. Copilot and Office funnel enterprise demand. Revenue recognition across the two companies remains opaque from the outside, but the strategic reality is stark: Microsoft’s bargaining power rises as OpenAI’s compute dependence grows. That can be a feature if the partner underwrites growth through generous credits, marketing muscle, and channel access. It is a risk if economics tilt toward the platform provider and away from the model maker. Expect investors to press for clarity on related-party terms, exclusivity provisions, and who captures which dollar of enterprise AI spend when services are delivered through Azure. For MSFT shareholders, the calculus is simpler. More OpenAI inference means more Azure consumption, even if OpenAI’s standalone margins compress. If the IPO locks in fresh cash for model training, Microsoft’s AI capex cycle stays well supported.
The blueprint for a blockbuster debut exists. SpaceX’s IPO jumped roughly 19 percent on day one, a reminder that public markets will pay for real technology moats, visible demand, and hard revenue. OpenAI can make a similar case if it delivers specifics. Investors will want a revenue run-rate with breakout between consumer subscriptions and enterprise licenses, gross margin trendlines that factor in compute, and a roadmap for lowering per-inference cost via model efficiency, scheduling, and potential silicon strategies. They will want contract summaries that lock in GPU supply at non-punitive pricing, plus visibility into power and data center partnerships. Governance will matter as much as growth: clear board oversight, risk controls around model safety, and a compensation structure aligned with long-term profitability, not just usage milestones. Miss on those disclosures and the market will treat the stock as a perpetual venture round with a ticker.
Alphabet’s (GOOGL) Gemini, Meta’s (META) open-source push, and focused challengers like Anthropic are narrowing the feature gap with rapid iterations and pricing tactics that pressure incumbents. On the chip side, Nvidia’s dominance keeps cost curves elevated while alternatives mature. Every quarter that inference remains expensive is another quarter that model providers fight for gross margin. Content licensing is a parallel pressure point. As publishers, labels, and creators negotiate usage and compensation, liability lines are being drawn in real time. The result is a market where price, performance, and legal clarity all move together. Any one of those vectors turning against OpenAI will make the $34 billion spend look heavier; if all three align, the number starts to look like table stakes for leadership.
The S-1, when it lands, should be read less like a prospectus and more like a cost-engineering brief. Key items to watch: the duration and magnitude of compute purchase commitments; related-party details with Microsoft, including termination rights and price-adjustment clauses; safety and trust spend as a percentage of revenue; and the scope of any ongoing governmental inquiries, including potential remedial actions. Investors should also expect a map of product-level unit economics, not just consolidated numbers, to separate consumer subscription margins from enterprise deployment margins. Finally, pay attention to churn, SLA credits, and support costs. In AI infrastructure, customer success is not a footnote; it is a line item that can rewrite the margin structure.
OpenAI’s spend sets the pace for a sector where the winners will be those who convert compute into defensible cash flow. For Microsoft, deeper OpenAI usage is a tailwind to Azure even as it amplifies scrutiny of partner economics. For Nvidia, the message is more demand for the foreseeable future and little near-term relief from customers still training and serving large models at scale. For the IPO itself, the path to a premium multiple runs through efficiency, transparency, and regulatory durability. The market has shown it will pay for audacity if the numbers square. The next few filings will tell whether $34 billion bought dominance or just bought time.
