Sora dies. The compute line is drawn.

The Thesis

The AI industry just drew a hard line between what is economically viable at scale and what is not, and that line runs straight through the compute budget.

The Signal

Three moves worth watching this week

1. OpenAI shut down Sora.

OpenAI shut down Sora on March 24, walking away from its standalone video app, its Disney character licensing deal, and a reported $1 billion Disney investment that never closed (TechCrunch, Variety, CNN).

The economics were never close: inference costs ran an estimated $15 million per day at peak while lifetime in-app revenue totaled $2.1 million [Appfigures via TechCrunch].

This is not a product pivot. It is the market correcting a false assumption that compute-intensive media generation can sustain consumer pricing. Every company budgeting for AI video should reassess which generation modalities survive contact with real unit economics.

2. MCP crossed 97 million monthly SDK downloads.

Anthropic's Model Context Protocol crossed 97 million monthly SDK downloads in March, just 16 months after launch, and is now governed by the Agentic AI Foundation under the Linux Foundation with OpenAI, Google, Microsoft, and AWS as members (Anthropic, WorkOS).

For context, React took roughly three years to reach comparable npm download volume. MCP is no longer an emerging option. It is the default integration layer for agentic systems, and teams building agent architectures without MCP compatibility are now building against the standard rather than alongside it.

3. NVIDIA launched its Agent Toolkit at GTC 2026.

NVIDIA used GTC 2026 to launch its Agent Toolkit, anchored by NemoClaw, a secure enterprise runtime for autonomous AI agents, with adoption commitments from Adobe, Atlassian, Salesforce, ServiceNow, and more than a dozen other major enterprise platforms (NVIDIA, Bain & Company).

Jensen Huang projected at least $1 trillion in compute revenue opportunity through 2027 and declared agentic AI at an "inflection point" (CNBC). The signal beneath the hardware announcements is architectural: NVIDIA is positioning agent security and orchestration, not GPU sales alone, as its enterprise growth layer.

The Playbook

The Compute Viability Stress Test: 5 questions to ask before committing budget to any AI-powered product or workflow

1. What is the per-unit inference cost at current usage, and at 10x usage? If the cost curve steepens rather than flattens with scale, the product has a Sora problem.

2. Does the workflow generate revenue or reduce cost at a rate that exceeds its inference spend? Sora generated $2.1 million against billions in compute. Productivity tools with measurable output gains pass this test. Creative generation tools often do not.

3. Is the AI modality text, structured data, or code (low inference cost) or image, audio, or video (high inference cost)? Prioritize investment in the former. Treat the latter as experimental until pricing proves otherwise.

4. Does the product depend on a single provider's proprietary model, or does it use a standard integration layer like MCP? Vendor lock-in is now also compute lock-in. MCP-compatible architectures allow model swapping as economics shift.

5. Is there a defined production governance layer, or is the deployment still demo-grade? If agent deployments lack runtime sandboxing, access controls, and audit trails, they are pilots, not products.

The Verification Test

Claim: Enterprise agentic AI has moved from pilot to production and is ready for scaled deployment.

Test: Ask the vendor for the number of agents running continuously in production environments with real customer or operational data (not sandboxed demos), the mean time between failures requiring human intervention, and the governance audit trail for agent actions.

Pass criteria: Named enterprise customers confirm agents running autonomously on production workloads for 90+ days with documented error rates, rollback procedures, and compliance audit logs.

Fail smell: Vendor cites "partnerships" and "integrations" with major platforms but cannot name a single customer running agents on live data without continuous human oversight, or governance consists of a policy document rather than an enforced runtime layer.

Links Worth Your Time

1. Bain on GTC 2026 — Argues the durable enterprise AI investments are infrastructure, governance, and security, not custom models. The build-vs-buy framing is shifting fast.
2. WorkOS on MCP in 2026 — The most complete technical overview of MCP's current state, from auth gaps to enterprise readiness. Essential reading for any team evaluating agent architectures.
3. Slate on the Sora shutdown — Goes beyond the product obituary to examine what the failure signals about OpenAI's financial position ahead of a potential IPO.
4. Fortune on the AI training gap — AI spending rising 44% in 2026 while workforce training budgets grow 5%. The gap between tool deployment and human readiness is widening.
5. Georgetown CSET on the White House AI Framework — Careful analysis of what the Framework actually proposes, where it is vague, and what Congress is likely to act on first.

The Builders & Doers Podcast

Chris Brisson on building Salesmsg, why texting is still the most underused channel in sales, and how AI agents are reshaping lead qualification. Covers accidental entrepreneurship, reinventing a stagnant business, and why distribution beats product for most founders.

Sources

1. TechCrunch. "OpenAI's Sora was the creepiest app on your phone, now it's shutting down." March 24, 2026.
2. Variety. "OpenAI Will Shut Down Sora Video App; Disney Drops Plans for $1 Billion Investment." March 24, 2026.
3. CNN. "OpenAI is shutting down its Sora video app." March 24, 2026.
4. Slate. "OpenAI's shock move with Sora should make you nervous about the economy." March 25, 2026.
5. Anthropic. "Donating the Model Context Protocol." 2026.
6. WorkOS. "Everything your team needs to know about MCP in 2026." March 2026.
7. Digital Applied. "MCP Hits 97M Downloads." March 2026.
8. NVIDIA Newsroom. "NVIDIA Ignites the Next Industrial Revolution." GTC 2026.
9. CNBC. "Nvidia GTC 2026: Agentic AI takes center stage." March 20, 2026.
10. Bain & Company. "Nvidia GTC 2026: AI Becomes the Operating Layer." March 2026.
11. CNN. "Judge blocks Pentagon's effort to punish Anthropic." March 26, 2026.
12. HBR. "AI Doesn't Reduce Work, It Intensifies It." February 2026.
13. AP/ClickOrlando. "AI's arrival complicates Big Tech climate goals." March 27, 2026.
14. Georgetown CSET. "Unpacking the White House National Policy Framework for AI." March 26, 2026.
15. Fortune. "Companies are pouring billions into AI and cutting training budgets." March 17, 2026.
16. White House. "National Policy Framework for Artificial Intelligence." March 20, 2026.

The Searchlight is a weekly field memo. It is not investment advice. Views are the editor's own.