Five days from now, on April 26, OpenAI will turn off the Sora app. The API limps along until September 24, and then it's gone. A product that was supposed to be the future of generative video, the wedge into Hollywood, the partner for a billion-dollar Disney deal, will be a line in the company's post-mortem docs.
Three senior OpenAI executives walked out the door on April 17 — Kevin Weil, Bill Peebles (the researcher behind Sora), and Srinivas Narayanan, the CTO of Enterprise Applications. On the same weekend, Cursor opened talks for a $2 billion funding round, Anthropic shipped Claude Design to compete directly with Figma-style workflows, and OpenAI's own Applications CEO told staff that Anthropic was a "wake-up call."
If you only read the headlines, this looks like another week of AI industry churn. If you're the person at your company actually responsible for AI roadmaps, contracts, and deployment, this week just handed you a homework assignment you cannot ignore.
Let me explain what actually happened, the math that forced the shutdown, and the four things I'm changing on my own roadmap at Zscaler because of it.
The Math Nobody Inside OpenAI Could Keep Ignoring
Sora was not a technical failure. It was a financial one, and the numbers are genuinely shocking when you put them side by side.
- Daily inference cost at peak usage: $15 million
- Projected annual cost: ~$5.4 billion
- Total lifetime revenue: $2.1 million
- Per 10-second video generation cost: approximately $1.30 in GPU time
- Peak downloads (Nov 2025): 3.33 million
- January 2026 downloads: 1.2 million — a 45% drop in one month
- January 2026 revenue: $367,000, down 32% from December
- Active global users: collapsed from roughly 1 million to under 500,000
Read those numbers again. A business burning five-point-four-billion-dollars-a-year was generating monthly revenue that wouldn't cover a mid-sized engineering team's salaries. That's not a cash-flow problem. That's a category error. And it had an enterprise cost, too: the Disney partnership — a $1 billion investment plus a three-year character licensing deal, signed in December 2025 — terminated before the companies ever announced it publicly. "No money was ever exchanged," according to the reporting.
OpenAI's Applications CEO said the quiet part out loud internally: the company had been "spreading its energy across too many applications." What he meant, reading between the lines, is that Anthropic's Claude Code is now doing roughly $19 billion in annualized revenue, with 80% of it coming from enterprises. While OpenAI was shipping a creative video model that couldn't pay rent, Anthropic quietly built a product enterprise CIOs were writing $1M+ checks for. Futurum's 1H 2026 decision-maker survey still shows 61% of organizations citing OpenAI GPT as their primary GenAI platform — but "primary" is a lagging indicator. The contract renewals happen later.
The Sora shutdown is what happens when a board, a CFO, and a pre-IPO capital structure look at those two realities and pick one.
The Pivot Has a Shape
OpenAI's "next phase of enterprise AI" framing — the company's exact language this week — is not marketing. It is a resource-allocation announcement dressed as a roadmap post. Three signals to pay attention to:
1. Senior leadership consolidated around enterprise. Weil led the Science initiative; that group is being folded into other research. Peebles was the face of Sora; his departure removes the internal champion for the creative flagship. Narayanan, the Enterprise Apps CTO, left for family reasons — benign on its face, but a CTO-level exit during a strategy pivot is never just about family. The org chart is being rebuilt around a "superapp" for enterprise productivity and a coding assistant meant to close the gap with Claude Code.
2. Compute is being redirected. Sora freed up an enormous GPU pool. That pool is going to the "Spud" LLM for enterprise productivity, and to the coding workloads that are now the commercial center of the company. The UK and Texas infrastructure scale-backs I've written about before make more sense in this light. OpenAI is not short on ambition; it is rationing compute the way a late-stage company rations engineering hours.
3. The IPO is driving the timeline. OpenAI's CEO wants a Q4 2026 IPO. The CFO has raised concerns about timing. Every decision coming out of the company right now is being filtered through "what does an S-1 reader want to see?" A $5.4B/year creative video burn does not survive that filter.
If you're an enterprise buyer, this is not neutral news. It means your vendor is in the middle of a capital-allocation pivot that will shape what they build, what they sunset, and what they integrate over the next 18 months. That pivot is happening whether your renewal calendar is ready for it or not.
The Vendor Durability Problem Just Got a Price Tag
I've been talking about vendor durability as an abstract risk for a year. This week it got concrete. If you were one of the enterprise customers who built on the Sora API, here's what your calendar looks like:
- April 26, 2026: App experience dies. Any workflow depending on the web or mobile product breaks.
- September 24, 2026: API sunsets. Any programmatic integration — marketing asset generation, training video automation, compliance explainer videos — needs to be migrated.
- Total migration window: Roughly five months.
Five months to rearchitect around a different provider, rewrite prompt pipelines, revalidate compliance, retrain users, and explain the change to the finance committee that approved the original program. For a mid-sized deployment, that's a reasonable quarter-and-a-half of engineering time burned on a migration nobody on your team asked for. For a regulated industry, add compliance revalidation, data residency review, and probably a security architecture pass.
And here's the uncomfortable piece: this is the best-case deprecation scenario. OpenAI told its enterprise customers in advance. They gave a multi-month window. The story plays out very differently when a vendor you depend on pulls a product with weeks of notice, or gets acquired, or runs out of money in a downturn.
Futurum's take lines up with what I'm telling my own team: "winners won't just be those with the best models, but those who can guarantee continuity." For the rest of us, the risk is no longer theoretical. It has a sunset date.
Four Things I'm Changing on My Roadmap at Zscaler
I run AI engineering for sales, marketing, finance, customer support, HR, and security at Zscaler. The Sora shutdown doesn't directly affect us — we never deployed video generation at production scale. But the shape of the story affects every architecture decision I'm about to sign off on. Four things changed in my planning this week.
1. Deprecation-safety is now a contract line item, not a wish. I've added explicit language to our AI vendor contracts requiring a minimum 12-month deprecation notice, bulk data export in an open format, and a meaningful continuity commitment for any API we depend on. If a vendor won't sign that, I price the risk into the contract or I walk. The Sora customers had five months. I'd rather not renegotiate under sunset pressure.
2. Every production integration gets an "abstraction floor." For any AI capability in a production workflow — summarization, code generation, content creation, agentic actions — the team now has to show me the switching cost to move to a different provider. If it's more than two engineer-weeks, we refactor until it isn't. This is the lesson from the Futurum data: single-platform dependencies are the fragility, and the remedy is architectural, not contractual.
3. We prefer vendors whose financial logic aligns with enterprise use. Anthropic's $30B ARR, 80% enterprise, four-times-cheaper training cost profile tells me they are not going to wake up one morning and torch an enterprise product to chase a consumer moonshot. OpenAI's cash burn, two roadmap pivots in six months, and IPO pressure tell me the opposite. I am not boycotting OpenAI — Zchat runs across Azure OpenAI and Gemini on purpose. But when I'm choosing which provider a new capability depends on, the financial durability of the counterparty is now a first-class criterion, not a footnote.
4. Multi-platform is the default, not the exception. Our AI Guard layer and the MCP gateway pilot give us a technical substrate to route workloads across providers based on cost, latency, or policy. I'm accelerating that work. The Sora shutdown is a proof-of-concept for a risk we've always understood on paper. The response is to make sure no enterprise workflow inside Zscaler gets married to a single API surface. If a model goes away, the workflow shouldn't.
The Deeper Signal: The Flagship Era Is Ending
Here is the part I think is being under-discussed. The Sora shutdown is not just OpenAI's story. It is the end of a particular era in AI — the era of flagship consumer products designed to dazzle, subsidized by investor capital, that were supposed to be the front door for enterprise adoption.
That playbook is done. You can see it in three places at once this week:
- Anthropic, with $30B ARR and a Series G that doubled its million-dollar enterprise customers from 500 to 1,000 in under two months, is executing a pure enterprise-first strategy.
- Cursor — a coding assistant that does exactly one thing — is raising $2 billion to do that one thing better. No video model. No audio model. No consumer flywheel.
- Anthropic's Claude Design launch this week is targeted at founders and product managers, not TikTok users. Its competitors are Figma and Canva, not Sora.
The money is moving toward products that pay for themselves inside enterprises, not products that generate cultural moments and bleed GPU cycles. For enterprise AI leaders, that's actually good news. It means the vendors are finally being disciplined by the thing we care about most — repeatable, durable revenue from customers who sign contracts, not users who churn.
But it also raises the stakes on our own discipline. If the vendors are going to stop subsidizing research moonshots with enterprise revenue, the enterprise side of the ledger has to earn its own keep. That means every AI project inside a large company needs a P&L, a governance posture, and a deprecation plan. The days of "we'll figure out the ROI later" are closing the same way Sora is closing.
What I'd Tell a CIO This Week
If I had ten minutes with a peer CIO this week, I would say three things:
First: audit your Sora dependencies now, not in August. Even if you never explicitly adopted Sora, check your marketing agency, your internal training team, your L&D vendor, and any integration that touched OpenAI's video endpoints. The April 26 app shutdown will break workflows people forgot they built.
Second: re-read your OpenAI, Anthropic, and Microsoft AI contracts this quarter. Ask your legal team a specific question: if the vendor discontinues the product line we've built on, what is our recourse, our notice period, and our data portability? If the answer is weak, negotiate before the next renewal. You will have more leverage this quarter than next.
Third: reframe the vendor conversation from "which model is best" to "which vendor will be there in 24 months." Model quality gaps close quickly. Vendor durability does not. The providers who will be standing in 2028 are the ones already generating durable enterprise revenue today. Optimize for that, and the capability will follow.
The Takeaway
Here is my one-sentence summary, for the inbox-skimmers: OpenAI just proved that enterprise AI strategy now has a vendor-risk dimension measured in months, not years, and the CIOs who don't bake deprecation math into their architecture and contracts are going to pay for it the hard way.
Sora is not the last product that will disappear under financial pressure. The correct response is not to panic, and not to pick the "safe" vendor — there isn't one. The correct response is to build an AI architecture where no single product's sunset can take down a workflow that matters. Abstract the APIs. Contract for deprecation notice. Prefer vendors whose business model aligns with yours. Run more than one model in production on purpose.
The vendors are getting disciplined. It's time we did too.
Rajesh Beri is Head of AI Engineering at Zscaler, where he leads the team building AI infrastructure across sales, marketing, finance, customer support, HR, and security. Views are his own.
Want to calculate your own AI ROI? Try our AI ROI Calculator — takes 60 seconds and shows projected savings, payback period, and 3-year ROI.
