The Product Desk

The most consequential competitive development this week isn't a model release — it's a go-to-market architecture shift that redraws the enterprise AI landscape. OpenAI launched Frontier, an enterprise agent management platform, and simultaneously signed multiyear distribution partnerships with McKinsey, Accenture, BCG, and Capgemini. This is the Salesforce playbook from the 2010s: own the platform, let consulting firms own the implementation margin, and use their Fortune 500 relationships as your sales channel.

The market's reaction was immediate and surgical. On Monday February 23, IBM dropped 13%, Salesforce/ServiceNow/Snowflake each fell ~4%, Microsoft dropped 2%, and Oracle fell 4.5%. This wasn't a broad tech selloff — the Nasdaq was down only 1.2%. Investors are pricing in a specific thesis: OpenAI is building the next enterprise software layer.

OpenAI showed investors a slide with Salesforce, Workday, Adobe, and Atlassian revenues alongside its own — the message was unmistakable: those companies' revenues are OpenAI's TAM.

Anthropic is running the opposite playbook. Its Claude Cowork launch on February 24 deliberately integrates with DocuSign, LegalZoom, and Salesforce, framing AI as replacing workers, not software. The market got the message — Figma jumped 10%, Salesforce rose 4%, ServiceNow inched up 1.4% on the Anthropic news. Peter McCrory, Anthropic's head of economics, explicitly framed this as a labor displacement story, not a software displacement story.

The 'frenemy' dynamic is now the dominant pattern. Both OpenAI and Anthropic maintain active partnerships with the very companies they're positioning to displace. This is textbook platform strategy: use partnerships to gain distribution and learn enterprise workflows, then vertically integrate once you have enough capability and customer trust.

The CrowdStrike Case Study Is Your Template

A cybersecurity executive replaced a CrowdStrike product — costing up to hundreds of dollars per user per month — with a Torq AI agent powered by OpenAI and Anthropic models, saving over $100,000 annually. The agent connected to raw Microsoft login data and replicated the core feature: flagging suspicious logins, finding dormant accounts, and automatically locking them down. CrowdStrike's response — saying it lets AI agents connect to its software — is the classic incumbent pivot from 'compete' to 'complement.' But the revenue is already gone.

The critical nuance: OpenAI's own COO publicly stated that AI has 'not yet really seen AI penetrate enterprise business processes.' This creates a paradox — the market is repricing SaaS stocks on a threat that hasn't materialized at scale yet. You have a window, probably 12-18 months, before displacement moves from anecdote to trend.

The Pentagon threatening to designate Anthropic a 'supply chain risk' — a classification normally reserved for foreign cyber warfare threats — is creating a vendor risk bifurcation across the entire AI industry that most product teams haven't priced in.

Here's what happened: Defense Secretary Pete Hegseth summoned Anthropic CEO Dario Amodei to a meeting described as a 'shit-or-get-off-the-pot meeting' over a $200M classified AI contract. Claude is currently the only AI model on classified military networks and was reportedly used during the January Maduro raid. Anthropic insists Claude should remain off-limits for domestic surveillance and autonomous weapons. The Pentagon's response: comply or face designation as a supply chain risk.

A supply chain risk designation wouldn't just kill Anthropic's direct Pentagon contract — it would require every Pentagon contractor to stop using Anthropic's tools. Not just the DOD itself.

The competitive landscape has already split. OpenAI and xAI removed military usage restrictions to pursue classified contracts. Alphabet agreed to remove safeguards for unclassified military use. Anthropic is the lone holdout. White House AI czar David Sacks publicly derided Anthropic last fall for advocating AI regulations.

Why This Is a Product Decision, Not Just a Policy Story

If you're building on Claude APIs and serve any customer in the defense supply chain, this is an active platform risk event. But the implications extend further. The Pentagon's posture is explicitly designed to set precedent for parallel negotiations with OpenAI, Google, and xAI — all of whom have already agreed to remove safeguards for unclassified military use but haven't reached terms on classified systems. This means the entire AI vendor landscape is being reshaped by government procurement leverage.

Simultaneously, Anthropic faces pressure from the other direction. DeepSeek, Moonshot AI, and MiniMax created over 24,000 fake accounts and ran 16 million exchanges to systematically distill Claude's capabilities — targeting agentic reasoning, tool use, coding, and computer vision. MiniMax alone was responsible for 13 million exchanges. Anthropic is fighting a two-front war: government coercion on one side, industrial-scale IP theft on the other.

For PMs, this creates a new evaluation dimension: AI vendor political/regulatory risk. A company fighting on two fronts may need to make hard choices — invest in anti-distillation measures (costing R&D resources), soften its ethical stance to win government contracts (changing its brand and product decisions), or double down on commercial enterprise (potentially meaning pricing changes). Any of these outcomes could affect your integration.

The two most well-funded AI companies on the planet can't predict their own cost structure — and that's the most important number in AI right now for product teams.

OpenAI's actual gross margin for 2025 was 33%, down from 40% in 2024 and a stunning 13 percentage points below their own forecast of 46%. Inference costs quadrupled to $8.4 billion, overshooting their summer forecast of $6.6B by $1.8B. Anthropic tells a similar story: they anticipated 40% gross margins for 2025 (already 10 points below their earlier target), coming off a negative 94% gross margin in 2024.

Nearly half of OpenAI's $8.4B inference costs ($3.9B) went to serving 865M non-paying weekly users, while only 45.5M users (5%) actually pay.

The root cause isn't GPU prices — average compute rental costs actually fell throughout 2025. It's demand and product mix. Reasoning models consume far more compute than traditional LLMs. Sora video generation is dramatically more expensive than text queries. The viral GPT-4o Ghibli image trend created demand spikes forcing OpenAI to buy expensive spot instances. For any PM building AI features, this is critical: your cost model needs to account for modality, not just API calls.

The Supply Constraint Is Getting Worse Before It Gets Better

On the infrastructure side, $96 billion in data center projects were blocked or delayed in Q2 2025 alone due to local opposition. JLL reports just 1% data center vacancy. Over 200 regulatory bills were introduced across U.S. states in 2025, with 40 becoming law and 10 new moratorium proposals filed in the past month alone. The Stargate joint venture between Oracle, OpenAI, and SoftBank has effectively dissolved, with OpenAI falling short of its 10GW capacity target.

But the Hardware Cliff Changes Everything for Late 2026

Taalas, a 2.5-year-old hardware startup, built a chip with Llama 3.1 weights physically etched into silicon, achieving approximately 17,000 tokens per second — 8.5x faster than Groq (~600 t/s) and 8x faster than Cerebras (~2,000 t/s). The tradeoffs are real: aggressive quantization causes elevated hallucination rates, and the model weights are permanently fixed. But the directional signal matters.

Meanwhile, Meta's $100B+ AMD deal — structured with performance-based warrants giving Meta up to 160 million AMD shares (~10% of the company) — signals that Nvidia's AI compute monopoly is ending. First gigawatt deployment on custom MI450 architecture is expected H2 2026. Combined with ASML's EUV breakthrough enabling 50% more semiconductor production by end of decade, the long-term inference cost trajectory is sharply downward.

The encouraging buried signal: OpenAI's compute margin specifically for paying users improved from 35% (Jan 2024) to 52% (end 2024) to 70% (Oct 2025). The inference efficiency gains are real and compounding — the problem isn't that AI can't be profitable, it's that the current product mix is drowning good economics in free-tier subsidy.

Buried beneath the enterprise warfare headlines is a signal that may matter more for consumer-facing PMs: Shopify is now running merchant ads inside ChatGPT, placing it alongside Facebook, Instagram, Google, and Snap in its advertising network. Target has confirmed participation in the pilot. The model is performance-based — merchants only pay on purchase.

A conversational AI interface is now being treated as equivalent to social media feeds for product discovery and purchase intent. With 910 million weekly active users, ChatGPT's reach rivals Instagram's.

Think about the user behavior shift this enables: instead of searching Google for 'best running shoes,' a user asks ChatGPT for a recommendation, and Shopify merchants' products appear as sponsored results within the conversation. This collapses the entire marketing funnel — awareness, consideration, and purchase — into a single conversational interaction.

The Trust Tradeoff

OpenAI's proposed ad model integrates directly with brand catalogs and APIs for 'more accurate, real-time, context-aware suggestions.' On paper, this sounds reasonable. But the moment ChatGPT recommendations are influenced by ad revenue, the trust dynamic changes irreversibly. Benedict Evans notes that ChatGPT's first advertisers are already live, described as 'entirely experimental.' Meanwhile, Walmart's ad business ($6.4B) now exceeds Snap's ($5.9B) — retail media is already a proven model that AI interfaces could supercharge.

For products that use AI-powered recommendations, discovery, or decision support, this creates a clear differentiation opportunity: 'Our AI recommendations are never influenced by advertising.' It also means that any workflow where your users currently rely on ChatGPT for product research or vendor selection may become less reliable as ad incentives enter the picture.

The Pricing Segmentation Signal

OpenAI is also fragmenting its subscription pricing with a planned $100/month Pro Lite tier — validating that the gap between $20 (Plus) and $200 (Pro) was too wide. The specific mention that Pro Lite could accommodate Codex usage suggests OpenAI is using feature access — particularly coding capabilities — as the primary tier differentiator. This creates a four-tier market: casual ($20), power user ($100), unlimited ($200), and enterprise (Frontier). Perplexity's painful lesson reinforces the design principle: when you need to monetize heavy usage, add value at a new price point rather than removing value from an existing one.