Ben Pouladian

OpenAI is still talked about like a research lab that happens to have a wildly popular product. That framing is starting to break.

Three things are becoming simultaneously true:

1. Unit economics are improving (model-running costs are taking a smaller bite out of revenue).

2. Enterprise is moving from “pilot” to “budget line item.”

3. The go-to-market is changing the SaaS competitive landscape (partners decompose workflows; OpenAI provides the model + platform).

Below is a stitched narrative using Bloomberg/The Information reporting, FundaAI enterprise/SaaS analysis, and Gene Munster’s valuation framework.

Note: None of this is OpenAI guidance. It’s third-party reporting + modeling. Treat the numbers as directional and the framework as the real takeaway.

• • •

The “Code Red” Framework

Before diving into the numbers, some context on how OpenAI thinks about competition.

Sam Altman recently confirmed that OpenAI triggers “code red” 1-2 times per year—whenever competitors like Gemini 3 or DeepSeek land a hit. These aren’t panic moments—they’re treated as 6-8 week sprints to fix product weaknesses.

Gemini 3’s launch in December was one of those moments. Google’s model briefly took the #1 spot on LMArena, and OpenAI’s US DAU dropped about 5% in the first week.

But here’s what struck me about Altman’s mindset:

Altman’s Code Red Philosophy:

➧ Paranoia is necessary to win

➧ Competitive shocks reveal weak spots

➧ The goal is to widen the lead, not just defend it

At the rate Google is shipping, those code reds might become monthly calendar invites next year.

The recovery in week two—DAU bounced back—suggests OpenAI’s user base has stickiness that survives competitive volleys. This is the posture of a company that believes it’s building durable advantages, not just riding a temporary lead.

The other Altman quote worth holding onto: if they had more compute, they could do so many more things. That’s crucial context for understanding the infrastructure commitments below. OpenAI isn’t capacity-constrained by market demand—they’re constrained by how fast they can build.

When I was scaling Deco Lighting from startup to over $100 million in revenue, I learned that being demand-constrained is a fundamentally different problem than being supply-constrained. OpenAI has the latter. That’s the better problem to have.

• • •

1. The Metric That Matters: Compute Margin Is Rising Fast

Bloomberg (citing The Information) reports OpenAI’s compute margin—defined as revenue after model-running costs—hit ~68% in October 2025, up from ~37% in January 2024. The path wasn’t a straight line—margins spiked to ~65% mid-2024, dipped back to ~50% around October 2024, then climbed again through 2025.

That one sentence is a big deal because it attacks the core bear case:

“LLMs are a marginless commodity because inference eats all the revenue.”

If compute margin is really moving from ~37% → 68% in ~20 months (with some volatility along the way), the story shifts from “AI is expensive” to “AI is getting cheaper faster than demand is getting bigger.”

What’s actually driving compute margin higher? A few non-controversial levers:

• Model routing (cheap models for easy queries, frontier models only when needed)
• Caching + reuse (especially for enterprise workflows)
• Inference optimizations (kernels, quantization, speculative decoding)
• Product/pricing mix shifting toward enterprise commitments

None of these require magical breakthroughs. They require operational competence and scale.

I spent time in Professor Fainman’s ultrafast nanoscale optics lab at UCSD working on silicon photonics. I understand what it means to optimize at the physical layer. What OpenAI is doing with inference isn’t revolutionary physics—it’s disciplined engineering. And disciplined engineering compounds.

🐻 The Bear Case on Margins

My friend Ram Ahluwalia pushes back hard on the “improved margins” narrative:

“Gross margin excludes the dominant economic cost of the product. Model training costs are the most significant cost to OpenAI—those are ‘below the line’ as Research Cost, not Gross Margin.”

— @ramahluwalia

Ram’s argument has teeth:

• Training vs. inference: Compute margin only captures inference costs. The massive training runs (GPT-5, etc.) sit in R&D—below the gross margin line.
• Model obsolescence: Unlike traditional software, LLMs depreciate fast. GPT-5.1 rendered 5.0 obsolete. Training costs are recurring, not one-time.
• Lease accounting: Operating leases with datacenter companies like CoreWeave convert CapEx to rental payments—flattering margins on paper.

The bull response: All true—but it’s also true of every frontier AI lab. The relevant question is whether OpenAI’s relative efficiency is improving faster than competitors. If they’re getting more inference per training dollar, and training costs are spreading across larger revenue bases, the unit economics still bend favorably. But Ram’s right that “compute margin” isn’t the whole story.

• • •

2. Demand Isn’t Just Consumer Subscriptions Anymore

FundaAI’s monitoring suggests consumer paid growth has decelerated recently (paid subs +5.5% MoM in Nov vs +7.5% in Oct), but argues that enterprise progress can keep ARR growth near ~10%.

This is the pattern you’d expect from a platform that’s graduating:

• Consumer growth normalizes
• Enterprise ramps with higher ARPU + longer retention cycles

💡 Key Insight

Enterprise monetization compresses time. It turns “usage” into “budget.”

The math worth walking through:

• • •

3. How OpenAI “Eats SaaS”: Workflow Decomposition + Unstructured Data

Here’s the key shift FundaAI is pointing at:

• ~70% of OpenAI use cases involve unstructured data (docs, emails, PDFs, call logs, knowledge bases).
• SaaS vendors historically win with structured workflows + structured fields inside their product.
• The AI opportunity is to treat the enterprise as a messy pile of data + processes and then orchestrate outcomes across systems.

The “SaaS moat” argument has been: vendors understand workflows + have domain know-how + own domain data.

FundaAI argues that dynamic is changing because PwC/Accenture/SoftBank-style partnerships can decompose workflows into stepwise agentic use cases, and mid-training can generalize across vertical scenarios.

One vivid example: “Mercury Project”—OpenAI paying 100 bankers $150/hour for data annotation.

From my Broadview investment banking days, I can tell you: that’s exactly right. The value in financial services isn’t generic intelligence—it’s knowing how to structure a pitchbook, how to model covenant packages, how to present comps to a skeptical board. OpenAI is building that expertise directly into the model.

The GTM Wedge

FundaAI describes a division of labor where OpenAI uses a small team (3 people: solution architect + security + model expert) for solution design (~3 weeks), while partners handle deployment and optimization.

That’s a clean scaling model: platform + channel partners.

Speed Matters

FundaAI notes Salesforce’s Agentforce deployments require ~6 months minimum, while PwC’s OpenAI-based solutions deploy within ~3 months.

In enterprise, time-to-value is frequently the deciding factor.

• • •

4. The SaaS TAM Is Enormous—and Budgets Won’t Magically Expand

FundaAI ’s SaaS sizing puts 2024 total market at ~$313B, split roughly half seat-based (~$164B) and half consumption-based (~$149B).

Here’s the important constraint:

⚠️ Reality check: Don’t assume IT budgets jump dramatically next year; hardware spend has already been squeezing software budgets.

And yet, the report frames next year as “Year One” of direct OpenAI/Anthropic competition with software—because if combined enterprise ARR reaches ~10% of the seat-based market, somebody is going to feel it.

🎯 Bottom Line

AI spend is not purely incremental. It reallocates budgets.

• • •

5. The Long-Range Model: Revenue Scales, Inference Costs Fall

FundaAI includes aggressive long-range projections. I’ve visualized the key dynamics:

My blunt take: This model is extremely optimistic on top-line (getting to $200B+ revenue by 2030 is not a base case).

But it’s useful because it shows what has to be true for OpenAI to justify “platform-scale” valuation:

• Inference cost share keeps dropping
• Enterprise/agents drive high-ARPU growth
• OPEX (especially research compute) stops scaling linearly with revenue

• • •

6. Valuation Lens: Why $830B Can Be Argued

Gene Munster highlights reporting that OpenAI is raising at a post-money valuation of ~$830B.

$830B

Reported Valuation Target

~24× projected 2026 revenue

Munster’s base revenue path estimate:

At ~$830B, that’s about 24× 2026 revenue, which Munster argues is still undervalued given platform potential.

But the same write-up emphasizes the uncomfortable part: OpenAI may rack up $100–$150B of cumulative losses from 2025–2029.

So the bet is clean:

🐂 Bull Case

“AI becomes the next cloud platform,” with improving compute margins making unit economics work.

🐻 Bear Case

“Capex and competition destroy returns,” and OpenAI becomes a perpetual compute buyer with limited pricing power.

The “code red” framework gives me some comfort here. A company that welcomes competitive pressure—that views Gemini 3 as healthy motivation rather than existential threat—has the right posture for a long fight.

• • •

7. What I’d Watch Next

If you want to track whether OpenAI is becoming a durable platform business vs a hype cycle:

1. Compute margin / inference cost share
If it keeps improving, the whole valuation debate changes.

2. Enterprise share of ARR
FundaAI’s thesis is enterprise becomes larger even if consumer growth slows.

3. Agent revenue traction
Agents are where “AI replaces SaaS workflows” becomes real.

4. Partner ecosystem depth
If partners can deploy faster and iterate without vendor lock-in, SaaS incumbents have a problem.

5. Budget reallocation signals
“Year One” means who starts missing renewals and which categories see pricing pressure first.

Closing Thought

OpenAI’s trajectory is turning into a classic platform story:

• Distribution (ChatGPT) creates demand
• Enterprise + agents create stickiness and ARPU
• Compute margin improvement turns growth into profit potential
• Partners turn a product into a workflow layer that can cannibalize SaaS

You don’t have to believe the most aggressive forecasts to see what’s happening: the economics are bending, and the enterprise wedge is real.

For investors focused solely on Google as the AI play, the opportunity cost question is worth considering. Google’s story is defensive—protecting search revenue, adding AI features to existing products. OpenAI’s story is offensive—building the AI-native layer that could capture value from every vertical it enters.

Both can win. But the upside profiles are different.

• • •

Resources

📊 Gene Munster: OpenAI at $830B Is Still Undervalued

📈 FundaAI: Deep|OpenAI Enterprise SaaS TAM Analysis

📰 The Information: OpenAI Compute Margin Report

🏢 OpenAI: 1 Million Business Customers

• • •

Ben Pouladian

CEO of BEP Holdings and publisher of BEP Research. Former co-founder of Deco Lighting (scaled to $50M+ revenue, exited 2019). Background in electrical engineering from UC San Diego, where I worked in Professor Fainman’s ultrafast nanoscale optics lab on silicon photonics. Chairman of the Leadership Board at Terasaki Institute for Biomedical Innovation.

🌐 benpouladian.com 𝕏 @benitoz