Rohit Prabhakar

I build agentic revenue systems for Fortune 50 companies

AI Weekly Memo – The Distribution Era Has Begun

by Leave a Comment

Week of May 18, 2026 | Signals from May 11 – May 17 For leaders who need signal, not noise.

The Thesis

Three weeks ago the bills came due for builders in the Reckoning Era. Two weeks ago for buyers in the Consumption Era. Last week AI stopped being sold and started being embedded in the Embedment Era. This week the question moved again. It is no longer about which model is best. It is about who owns the channel that model ships through – hence the beginning of the AI Distribution Era.

Anthropic put Claude Platform on AWS, making it the only frontier model available on all three major clouds, billed on a single invoice that retires against your existing AWS commitment. OpenAI put Codex on your phone, turning the device in your pocket into the control surface for autonomous agents running on machines you are not even sitting at. And China looked at cleared, export-approved Nvidia H200 chips and said no, choosing a slower domestic stack over dependence on someone else’s supply chain.

Three different stories. One pattern. The model is becoming a commodity. The distribution channel is becoming the moat. Whoever owns the channel owns the pricing power, owns the customer relationship, and owns the lock-in. Welcome to the Distribution Era. The strategic question for your board is no longer “which AI is best.” It is “who controls the pipe our AI flows through, and what does that cost us in five years.”

3 Questions for the Board This Week

  1. The Lock-In Question: If our AI consumption is billed through our cloud provider on a single invoice, who actually owns our switching costs, and have we modeled what that does to our negotiating leverage at renewal? (AWS)
  2. The Control Surface Question: As autonomous agents take on long-running work, who in our organization can start, stop, approve, and audit them, and is that control surface governed or improvised? (OpenAI)
  3. The Sovereignty Question: China just refused export-approved US chips to protect its own stack. Do we have a contingency plan if our AI supply chain bifurcates into a US stack and a China stack? (Reuters via WION)

The Signals: Why These Questions Matter Now

1. Claude Platform Landed on AWS. The Distributor Just Won.

The News: On May 11, Anthropic launched Claude Platform on AWS, two weeks after OpenAI put GPT-5.5 and GPT-5.4 on Amazon Bedrock (AWS, Caylent). AWS now hosts both frontier model families on a single bill, with authentication through AWS IAM, audit logging through CloudTrail, and consumption-based pricing that retires fully against existing AWS commitments. Setup takes about ten minutes. Claude Platform on AWS goes beyond Bedrock with managed agents, code execution, web search, and day-one access to new features through Anthropic’s native APIs. It launched in 18 AWS regions. Claude is now the only frontier model available on all three major clouds (AWS, Google Cloud, Azure); OpenAI is on two. Anthropic holds roughly 38% of token consumption on Bedrock and reached $30 billion ARR in April, passing OpenAI’s $25 billion while spending roughly four times less on training.

Strategic Insight: This looks like a buyer’s market. It is not. When you can swap Claude for GPT-5.5 with a one-line code change and the same invoice, the models become interchangeable and the platform hosting them owns the pricing power. This is the framing your board needs to hear: we have seen this movie in cable TV, app stores, and cloud computing itself. The distributor always wins. The dangerous part is subtle. Every A/B test you run between models on the same cloud teaches the cloud provider more about how to price your next contract than it teaches you about model quality. The convenience that makes adoption frictionless is the same convenience that erodes your negotiating leverage at renewal. Near-zero switching costs between models is not the same as near-zero switching costs away from the platform. Those are opposite things, and the platform is counting on you confusing them.

Board Reality: Procurement and the CIO need to separate two questions that feel identical and are not. Question one: can we switch models easily? Yes, and that is good. Question two: can we switch distribution channels easily? Increasingly no, and that is the risk that compounds. Model the five-year total cost of a single-invoice AI relationship the same way you would model a single-vendor ERP lock-in, because structurally that is what it is becoming. Negotiate exit terms now, while you still have the leverage of being early.

2. The Phone Became the Agent Control Surface

The News: On May 14, OpenAI brought Codex into the ChatGPT mobile app on iOS and Android, in preview, across every plan including Free (OpenAI, TechCrunch). More than 4 million people now use Codex weekly. The phone does not run the code. It becomes the control surface for Codex sessions running on a laptop, a Mac mini, or a managed remote environment, connected through a secure relay. From the phone you can review diffs, approve commands, switch models, redirect tasks, and monitor terminal output in real time. Files, credentials, and permissions stay on the host machine. Remote SSH went generally available, and HIPAA-compliant Codex shipped for eligible Enterprise workspaces. This follows Anthropic’s Claude Code Remote Control, which shipped the same capability in February. As one analysis put it, OpenAI did not invent mobile-connected agentic coding; Anthropic shipped it four months earlier. The race is now over who owns the supervision layer, not who can shrink a development environment onto a screen.

Strategic Insight: Knowledge work is becoming asynchronous agent supervision, and most organizations have no governance model for it. The new rhythm is concrete: an employee starts a task at their desk, walks away, approves the output from their phone over coffee while the agent has been working autonomously in between. This is not a productivity feature. It is a structural change in what work is. The implications cascade. If an agent runs for two hours unsupervised and an employee approves its output from a phone in 15 seconds, who is accountable for what that agent did? Where is the audit trail? What stops an approved-on-mobile action from touching production? The Codex architecture keeps credentials on the host machine, which is the right design, but the human judgment has moved to a four-inch screen in a coffee shop, and that is where governance has to follow.

Board Reality: The CISO and Chief AI Officer need an agent supervision policy before this becomes ambient. Three questions define it. Who is authorized to approve autonomous agent actions, and from what devices? What classes of action require desk-based review versus mobile approval? Where is the immutable audit log that captures what the agent did, what the human approved, and the time gap between them? If the answer to any of these is “we have not decided,” you are already running ungoverned autonomous work, you just have not measured it yet.

3. China Refused Export-Approved Nvidia Chips. The Stack Is Bifurcating.

The News: This week President Trump said aboard Air Force One that China “chose not to” buy approved Nvidia H200 AI chips, preferring to develop domestic alternatives (WION, Tom’s Hardware). The US had cleared roughly 10 Chinese technology giants, including Alibaba, ByteDance, JD.com, and Tencent, to buy up to 75,000 H200 chips each through intermediaries like Lenovo and Foxconn. The US wanted 25% of the export revenue, and the arrangement required the hardware to physically pass through US territory for testing. Beijing’s customs authorities have blocked the imports, allowing only universities and R&D labs to acquire the chips. China has committed incentives reportedly worth up to $70 billion to support domestic chipmakers. Huawei’s Ascend roadmap runs 950PR in 2026, 960 in 2027, 970 in 2028, with its own high-bandwidth memory. Nvidia’s market share in China has fallen from 95% before sanctions to under 60%. US Trade Representative Jamieson Greer called the purchase decision a “sovereign decision” for China.

Strategic Insight: The single global AI stack is over. There are now two, and they are diverging on purpose. China is accepting a slower, less capable domestic stack today in exchange for not depending on a supply chain that another government can switch off. That is not an emotional decision. It is a strategic one, and any Fortune 500 with meaningful China exposure now faces the same calculation in reverse. The assumption that you can run one AI architecture, one model strategy, and one compute supply chain globally is no longer safe. The bifurcation is not coming. It is here, it is policy-driven on both sides, and it will widen.

Board Reality: Any company with China operations, China revenue, or a China-touching supply chain needs a two-stack contingency plan: a US-aligned AI stack and a China-aligned AI stack, with explicit decisions about data, models, and compute in each. This is no longer a 2028 scenario-planning exercise. It is a this-year architecture decision, and the companies that make it deliberately will outperform the ones that have it forced on them.

3 Strategic Actions for This Week

  1. Run the Distribution Lock-In Model. CFO and CIO co-own. Model the five-year total cost of your single-invoice AI relationship the way you would model single-vendor ERP lock-in. Separate “can we switch models” (good, keep) from “can we switch channels” (the compounding risk). Negotiate exit terms now while early-adopter leverage still exists.
  2. Write the Agent Supervision Policy. CISO and Chief AI Officer. Define who can approve autonomous agent actions, from which devices, for which classes of action, and where the immutable audit log lives. If autonomous work is already happening ungoverned, this is overdue, not premature.
  3. Build the Two-Stack Contingency. CIO, General Counsel, and Chief Strategy Officer. If you have China exposure, design the US-aligned and China-aligned AI stacks explicitly, with data, model, and compute decisions made deliberately rather than reactively.

Bottom Line

Three weeks ago the bills came due for builders in the Reckoning Era. Two weeks ago for buyers in the Consumption Era. Last week AI moved inside the workflow in the Embedment Era. This week the lesson is sharper and older than AI itself: the company that controls distribution controls the economics.

Anthropic made Claude the only model on all three clouds and put it on the same bill you already pay. OpenAI made your phone the place you approve work an agent did while you were not watching. China decided that controlling its own stack was worth more than the best available chips. Different stories, one truth. The model layer is commoditizing. The distribution layer is consolidating. Power is moving from what the AI can do to who controls how it reaches you.

If your board is still debating which model is best, you are optimizing the layer that is becoming free while ignoring the layer that is becoming the moat. The Distribution Era is here. The question is whether you own your channel, or someone else owns you through it.

This memo is part of the Market-of-One framework. Subscribe to the Weekly AI Memo for the board-level read every week.

Disclaimer: AI used for content and creative