The Product Desk

The Distribution Event

A user upgrades to iOS 27, iPadOS 27, or macOS 27 and is asked to pick an AI provider. She picks once. From that moment, every system-wide AI request — summarize this, draft a reply, generate this image — routes to whichever third-party AI model she selected for text generation, image generation, and editing. The named beneficiaries are Google and Anthropic, which tells you the bar is foundation-model scale. Every system-wide AI query goes through one chosen provider.

There is no browsing. Users pick once. Whoever is in the default slot on day one gets the query volume. Whoever is not, does not.

Why This Is Different From Previous Platform Shifts

The $250M Siri settlement is the backstory. Apple promoted AI features "that did not exist at the time, do not exist now, and will not exist for two or more years." It could not ship competitive AI fast enough, so it became the selector layer instead. This is the App Store playbook applied to AI: control distribution, extract rents, let others compete on capability.

OpenAI is running the same play from the opposite direction — a dedicated AI phone with dual-NPU architecture targeting 30M units in 2027-2028. Google's Remy agent is in internal dogfood, potentially debuting at I/O May 19-20. Meta has an agentic assistant targeting pre-Q4 2026. Four companies are building the next interaction layer simultaneously.

The Three-Cell Diagnostic

If the product is a model: integration work starts now. Latency budgets on-device, Apple's review process, being on the partner list before it goes public. That is a real distribution win at billion-device scale.

If the product is a model wrapper: the default-assistant slot is where distribution goes to die. The user picks the underlying model. The wrapper layer disappears.

If the product is an application built on a model: the pricing conversation with the model vendor changes in two quarters, because their marginal query cost is about to drop. The core product is safe only if it owns data or workflow the assistant cannot replicate.

The Forcing Function

Pull the last 30 days of session data. Separate sessions where user intent formed elsewhere (the user arrives with a question) from sessions where intent forms inside the product (the user discovers what to do next). If the first bucket is larger, the iOS 27 moves hit the roadmap this quarter. If the second is larger, the product has a workflow moat the selector layer cannot intermediate.

Three Cost Curves Moving Simultaneously

Three infrastructure developments landed this week that collectively invalidate Q1 cost assumptions for AI features:

1. SubQ's 12M-token context window at 1/1000th the compute of frontier models — $8 per benchmark run versus ~$2,600 on Opus. RULER 128K score of 95%, competitive with Opus 4.6 and DeepSeek V4 Pro.
2. Vision agents cost 45x structured API paths — and better models don't close the gap because screenshot volume (not accuracy) drives token consumption.
3. Multi-Token Prediction drafters at 78M parameters deliver 3x throughput with no quality degradation, with day-0 support across vLLM, SGLang, Ollama, and llama.cpp.

A feature that cost $0.03 per interaction six months ago now costs $0.012 on optimized routing. At volume, that's the line between a product and a subsidy.

What the Context Cost Cliff Means for RAG

SubQ's numbers are self-reported and need independent validation. But the useful question isn't whether these specific numbers hold — it's what happens to chunking strategies, multi-hop retrieval, summarization chains, and progressive document disclosure when context drops 100x in price. All of those patterns exist because context was scarce. About 40% of most RAG roadmaps becomes unnecessary if the pricing holds. That is not the same as RAG being dead — retrieval still solves freshness and access control — but it's a different product with a different cost profile.

The Vision Agent Trap

Teams built vision-first because prototyping was easier. The cost curve is flat against model improvements — a model that is 50% more accurate still sends thousands of input tokens per screenshot. The forcing function is a simple tag on your automation backlog:

The Inference Router Opportunity

DigitalOcean's Inference Router delivered 61% cost reduction through intelligent model routing (selecting optimal models by cost, latency, and quality per request). Combined with MTP drafters and the SubQ context breakthrough, the features shelved because 'inference is too expensive at scale' deserve a re-run through current economics. Some are shippable now.

The Specific Threat

A pricing PM at a retail company opened her model's feature importance dashboard this week and saw that device type and browsing history were two of the top five signals. That is the problem. Maryland's Protection From Predatory Pricing Act takes effect October 1, 2026, the first US state-level ban on AI-powered dynamic pricing. Governor Wes Moore named the target directly: "predictive AI that determines what we need, when we need it, when we'll pay for it and when we'll pay more for it." That language describes every modern pricing optimization system that uses individual user signals.

Per the New York Times, roughly 33 additional states have similar bills in motion. Some target surveillance pricing. Some target surge pricing on essentials. A few cover any algorithm that varies price by user attribute. The drafts do not agree with each other, which makes a patchwork harder to comply with than a single ban.

The strictest state's rules become your de facto national standard because building 50 state-specific pricing engines is impractical. This is GDPR-for-pricing.

What's Banned vs. What Survives

The pitch is usually "personalization." What the pricing engine is actually doing is price discrimination using individual user signals: purchase history, device type, behavioral patterns that move the price person-to-person for the same product. The law does not touch contextual and temporal signals: store location, time of day, inventory levels, demand tiers applied uniformly.

Revenue optimization does not have to die. It has to become transparent and non-discriminatory:

• Time-based pricing (happy hour, early bird) likely survives
• Volume-based pricing and membership tiers likely survive
• Per-user price discrimination from behavioral signals does not survive

The Audit Framework

The 2x2 a pricing PM can run on Monday: on one axis, label every input as individual, contextual, or temporal. On the other, measure the revenue lift attributable to each group. Pull the individual signals out and rerun the model. If the lift collapses when those signals come out, the product was extracting, not personalizing. That is worth knowing before a regulator tells you.

Why This Is a PM Problem, Not a Legal Problem

Compliance work slips because nothing breaks on the demo. But re-architecting a pricing engine to be scoped by jurisdiction, category, and customer segment, with decision logs a regulator will accept, is roughly four quarters of engineering. Starting the work after October 1, 2026 means operating illegally during the build.