Odd Lots: Henry Blodget on the Problem for OpenAI (Podcast)

Bloomberg's Odd Lots podcast features Henry Blodget presenting a bearish thesis on OpenAI's long-term competitive position — arguing that despite building the most capable AI models, the company faces a fundamental distribution problem that may prevent it from capturing value proportional to its technology leadership. The core insight: enterprises don't want standalone AI platforms; they want AI embedded in the tools they already use. This dynamic transforms OpenAI from a direct enterprise vendor into a model supplier — powerful but commoditizable.

For investors, enterprise technology buyers, and finance professionals evaluating AI vendor strategy, Blodget's framework provides a critical lens: the company building the best AI model may not be the company that captures the most enterprise value from AI.

The Distribution Problem Detailed

Blodget's argument centers on a historical analogy and three contemporary observations:

The Intel analogy. Intel built the world's best processors for decades. But Intel never captured the software application revenue that its processors enabled. Microsoft (operating system), Oracle (databases), and thousands of application companies captured the value. Intel earned margin on chips but was perpetually commoditized by competitors (AMD, ARM) that offered "good enough" processing at lower cost.

OpenAI is in the same position: building the "processor" (GPT models) while Microsoft (Copilot), Salesforce (Einstein), and SAP (Joule) build the application layer that enterprises actually buy. These companies pay OpenAI for model access — but capture 5-10x more revenue per customer than OpenAI earns from the underlying API call.

Observation 1: Enterprise sales requires relationships, not technology. A Fortune 500 CFO doesn't evaluate AI models technically. They ask their existing vendor (Microsoft, Salesforce, SAP): "Does your product have AI?" The answer is "yes" — regardless of whether it's powered by OpenAI, Anthropic, Google, or open-source models. The model provider is invisible to the enterprise buyer.

OpenAI has 200 enterprise salespeople. Microsoft has 60,000. Salesforce has 30,000. In enterprise sales, distribution beats technology. The company with the customer relationship wins the deal — and that's rarely the model provider.

Observation 2: Switching costs favor incumbents, not OpenAI. An enterprise using Salesforce CRM with Einstein AI has: 5 years of customer data in Salesforce, thousands of customized workflows, trained users, and integration with 50+ other systems. Switching to "ChatGPT for CRM" requires replacing all of this. The switching cost is $10-50 million and 2-3 years — ensuring Salesforce retains the relationship regardless of whether OpenAI's model is marginally better.

Observation 3: Model quality is converging. GPT-4 was dramatically better than alternatives in 2023. By 2026, Claude 4, Gemini 2.5, Llama 4, and Mistral Large perform within 5-10% of GPT-5 on most enterprise tasks. This convergence means enterprises don't need the BEST model — they need a GOOD ENOUGH model embedded in their existing workflow. "Good enough" is available from every major platform vendor.

The Revenue Model Vulnerability

Blodget quantifies the problem:

ChatGPT consumer: $20/month × 5M paid users = $1.2B annual revenue. Growth plateauing as casual users find free tier sufficient and power users hit capability ceilings.

API/developer revenue: Usage-based pricing generating ~$2B annually. But: developers use whichever model is cheapest for their quality threshold. As Llama (free), Mistral (cheap), and Claude (competitive) improve, API pricing faces downward pressure.

ChatGPT Enterprise: $500K-$5M contracts targeting Fortune 500. This is OpenAI's growth bet — but Blodget argues it's failing because enterprises prefer Copilot ($30/user/month through existing Microsoft EA) over ChatGPT Enterprise (separate platform, separate training, separate integration).

The math problem: If enterprise prefers embedded AI (via incumbents) over standalone AI (OpenAI direct), then OpenAI's addressable market is the wholesale API layer — where it competes on price against increasingly capable alternatives. This is a lower-margin, higher-volume business that doesn't justify a $100B+ valuation.

The Bull Case (what Blodget's thesis may underestimate)

Consumer lock-in. 200M+ ChatGPT users with conversation history, custom GPTs, and established workflows may resist switching even if alternatives improve. ChatGPT has become "the interface" for AI in many users' mental models.

Developer ecosystem. OpenAI's API has the deepest developer ecosystem (most tutorials, most integrations, most third-party tools built on GPT). Ecosystem effects create stickiness beyond raw model quality.

Research velocity. OpenAI's investment in research ($3-4B annually) may maintain a capability gap that prevents convergence. Each GPT generation could re-open the quality gap that competitors spent 18 months closing.

Custom model moats. Enterprises that fine-tune GPT on proprietary data create switching costs with OpenAI specifically. These customizations don't transfer to competing models easily.

Why This Matters

For enterprise AI buyers: Don't evaluate OpenAI in isolation. Ask: "Can I get equivalent AI capability through my existing Microsoft/Salesforce/SAP relationship at lower total cost?" In most cases, the answer is yes — making standalone OpenAI contracts harder to justify unless the use case specifically requires GPT's unique capabilities.

For CFOs budgeting AI spending: The "Intel Inside" dynamic means AI capability becomes embedded in software you're already paying for (via price increases). Your AI budget may be invisible — hidden in 15-30% SaaS price hikes rather than a separate "AI platform" line item. Track this holistic AI cost.

For investors: OpenAI's $100B+ valuation requires enterprise revenue growth that Blodget argues won't materialize at projected levels. If the "Intel Inside" thesis is correct, OpenAI deserves a wholesale API company valuation (5-8x revenue) rather than a platform company valuation (20x+ revenue) — implying 50-70% downside risk.

For accounting firms: Client advisory on AI vendor selection should include: total cost of ownership analysis (embedded AI via existing vendor vs. standalone OpenAI + integration cost), switching cost modeling, and multi-vendor risk assessment that accounts for the model convergence trend.

For competing AI companies (Anthropic, Google, Mistral): The "Intel Inside" problem affects all model providers equally. The winners in enterprise AI may not be model builders at all — but the application companies (Microsoft, Salesforce, SAP) that embed whatever model is cheapest/best into their existing customer relationships.

The key takeaway: Blodget's "Intel Inside" thesis argues that OpenAI building the best AI model doesn't guarantee capturing enterprise value — because enterprises buy AI embedded in existing vendor platforms (Copilot, Einstein, Joule) rather than standalone AI services, potentially making OpenAI a wholesale model supplier valued at 5-8x revenue rather than the 20x+ its valuation implies.