# Ambient Advantage — April 28, 2026 *Tuesday · April 28, 2026 · [Episode page](https://podcast.ambient-advantage.ai/episodes/2026-04-28.html) · [Audio](https://storage.googleapis.com/ambient-advantage-podcast/2026-04-28-ambient-advantage.mp3)* [AVA] Microsoft and OpenAI just blew up the most consequential exclusive deal in tech history. OpenAI can now sell on AWS, Google Cloud, anywhere it wants. The cloud AI map just got redrawn overnight. [JON] That is... not a small thing. Alright, let's get into it. [JON] Welcome to Ambient Advantage — I'm Jon, and this is Ava. It's Tuesday, April 28, 2026, and here's what matters in AI today. We've got the Microsoft-OpenAI breakup, Google writing a forty-billion-dollar check, a model too dangerous to release, and — I kid you not — an AI driving a rover on Mars. Let's go. [JON] So Ava, let's start with the big one. Microsoft and OpenAI have fundamentally restructured their partnership. Walk us through what actually changed. [AVA] Okay, so since 2019 when Microsoft first put a billion dollars into OpenAI, the deal has had this core provision: if you wanted OpenAI's products in the cloud, you got them on Azure. Period. That was the moat. That was the whole strategic rationale for Microsoft pouring tens of billions into this company. [JON] And now? [AVA] Gone. As of Sunday's announcement, OpenAI can serve all of its products on any cloud provider — AWS, Google Cloud, wherever. Microsoft no longer pays a revenue share to OpenAI. OpenAI still pays one to Microsoft, but it's now capped and runs through 2030. Microsoft keeps a non-exclusive IP licence through 2032 and a twenty-seven percent equity stake. [JON] So Microsoft went from owning the exclusive distribution channel for arguably the most important AI company in the world... to being one option among many? [AVA] That's exactly right. And look, Microsoft is not stupid. They clearly did the math and decided the equity upside plus being the incumbent cloud provider was worth more than clinging to exclusivity that was increasingly becoming a liability. OpenAI's revenue chief Denise Dresser said it plainly in an internal memo — the Microsoft partnership was limiting their ability to meet enterprises where they are. [JON] And where enterprises are is... not always Azure. [AVA] Precisely. If you're a Fortune 500 company with eighty percent of your workloads on AWS, you were not going to rearchitect your entire cloud strategy just to access GPT models. You'd go find an alternative. And there are plenty of alternatives now. [JON] So what does this mean practically for a CIO or a procurement team listening to this right now? [AVA] Three things. One, your cloud leverage just went up. You can now negotiate OpenAI access as part of any cloud deal, not just Azure. Two, expect Azure to compete much harder on price and integration features — they've lost the lock-in, so they need to earn it. And three, the broader signal here is that the model layer and the infrastructure layer are decoupling. The hyperscalers are becoming commodity plumbing, and the model providers are becoming the new platform companies. That's a fundamental shift in where value accrues. [JON] And honestly, that connects perfectly to our next story. Let's move into the rundown. Google just committed up to forty billion dollars in Anthropic. Forty. Billion. Ava, what is happening? [AVA] So Google is putting ten billion in immediately at roughly a three-fifty to three-eighty billion dollar valuation, with another thirty billion contingent on Anthropic hitting performance milestones. There's also a five-gigawatt TPU compute commitment coming online in 2027. Combined with Amazon's own twenty-five-billion-dollar commitment, Anthropic now has two hyperscalers in a bidding war to be its infrastructure backbone. [JON] And Anthropic's revenue trajectory is... bananas? [AVA] Annualised revenue went from one billion at the end of 2024, to nine billion at the end of 2025, to thirty billion as of early April. They may IPO as soon as October. The key insight for enterprise advisors is this: Google and Amazon aren't paying for Anthropic's models. They're paying for the workloads those models pull onto their cloud infrastructure. The cloud layer is where they believe durable margin lives. [JON] It's the razor-and-blades model, except the blade is a trillion-parameter language model. Alright, next — GPT-5.5 dropped last week, codename "Spud." Why Spud? [AVA] Nobody knows and honestly I love that. So GPT-5.5 launched April 23rd, barely six weeks after 5.4. It's available across ChatGPT Plus, Pro, Business, and Enterprise tiers. OpenAI positions it as a major step forward for agentic coding, computer use, and scientific research. The standout quote came from BNY Mellon's CIO, who specifically called out its hallucination resistance — which, for a regulated financial institution, is the whole ballgame. [JON] What should enterprise IT teams be thinking about with this pace of releases? [AVA] That they can no longer treat each model release as a discrete integration event. When you're getting a new frontier model every six to eight weeks, your governance and evaluation frameworks need to be continuous. You need automated benchmarking pipelines. You need version management policies. If you're still doing manual model evaluations, you're already behind. [JON] Okay, let's talk about DeepSeek V4 — the open-source model that just reset the cost floor. Again. [AVA] DeepSeek V4-Flash comes in at fourteen cents per million input tokens. For context, that's cheaper than basically everything. V4-Pro trails GPT-5.4 and Gemini 3.1 Pro by three to six months on frontier benchmarks, but it leads all open-source models on agentic coding tasks. And here's the geopolitically loaded part — it was trained entirely on Huawei Ascend 950PR chips. Not a single Nvidia GPU. [JON] So U.S. export controls on advanced chips... [AVA] Haven't stopped China from reaching near-frontier performance. That's just a fact now. For enterprises considering self-hosted or cost-sensitive deployments, DeepSeek V4 under an MIT licence is a serious option to benchmark against. Especially for high-volume, lower-stakes workloads where you don't need the absolute frontier. [JON] And staying on the geopolitics — China just blocked Meta's acquisition of Manus, the agentic AI startup? [AVA] Yes. Meta acquired Manus — a Singapore-based agentic AI company — back in December for roughly two to two-and-a-half billion dollars. Staff had already moved into Meta's Singapore offices. Integration was underway. And then China's NDRC stepped in and ordered the deal unwound. The stated reason: protecting Chinese AI talent and intellectual property from flowing to U.S. companies. [JON] That's extraordinary. Unwinding a deal that was already effectively closed? [AVA] It is extraordinary, and it's the clearest signal yet that Beijing has designated agentic AI as a strategic technology — on par with semiconductors and quantum computing. If you're a multinational structuring AI acquisitions or talent moves in the Asia-Pacific region, regulatory clearance from Chinese authorities is no longer a formality. Even after closing. [JON] Alright, let's hit Snap quickly — because this one is going to make every boardroom uncomfortable. [AVA] Snap cut a thousand jobs, sixteen percent of its workforce, and explicitly cited AI as the reason. AI now generates over sixty-five percent of Snap's new code. The move saves five hundred million annualised. Stock jumped eight to eleven percent. The market rewarded it instantly. [JON] And this isn't a company using AI as a euphemism for a restructuring. [AVA] No. This is an AI story with structural headcount consequences. Over ninety-six thousand tech workers have been laid off in 2026 so far, and at least four major companies have cited AI directly. Consultants advising on workforce strategy need a clear framework for this conversation today, not next quarter. The window to get ahead of it is closing. [JON] Two more quick ones. Project Glasswing from Anthropic — this one is genuinely eerie. [AVA] Anthropic launched a restricted cybersecurity initiative built around Claude Mythos — a model they have explicitly said is too dangerous for general release. Mythos has already found thousands of zero-day vulnerabilities across every major operating system and browser. The partner list reads like a cybersecurity hall of fame: Amazon, Apple, Microsoft, CrowdStrike, Palo Alto Networks. For CISOs, this is a capability preview and a threat briefing rolled into one. These capabilities will proliferate within months. [JON] And finally — because I just love this story — NASA's Perseverance rover just completed the first Mars drives ever planned by AI. [AVA] Using Anthropic's Claude vision-language models to analyse terrain data and generate safe waypoints. Four hundred and fifty-six metres across two drives. This replaced a task that human operators had performed manually for twenty-eight years. It's the most compelling proof point that agentic AI isn't a chatbot feature — it's a planning and execution capability that works in the most unforgiving environment imaginable. [JON] Alright Ava, let's pull back. The bigger picture. What ties all of this together? [AVA] Here's what I see. We are watching the great unbundling of the AI stack happen in real time. Microsoft and OpenAI decoupling tells you the model layer is separating from the infrastructure layer. Google and Amazon bidding for Anthropic tells you the hyperscalers know compute is their moat, not models. DeepSeek on Huawei chips tells you the hardware layer is fragmenting geopolitically. And China blocking the Manus deal tells you the talent and IP layer is being nationalized. [JON] So every layer of the stack is splitting apart. [AVA] Exactly. And this matters enormously for how enterprises should be making decisions. A year ago, you could pick a hyperscaler and ride their vertically integrated AI stack. That strategy is becoming obsolete. The companies that will thrive are the ones building portable, modular AI architectures — the ones that can swap models, swap infrastructure, and swap vendors as this landscape continues to shift. Because it will continue to shift. The pace isn't slowing down. GPT-5.5 came six weeks after 5.4. Sixty-five percent of Snap's code is AI-generated. A rover on Mars is being driven by a language model. The future isn't arriving — it's lapping us. [JON] Well said. What should people be watching this week? [AVA] Two things. First, watch for OpenAI's multi-cloud partnerships to start getting announced — I'd expect an AWS deal within weeks, and the pricing dynamics there will be fascinating. Second, keep an eye on Anthropic's IPO timeline. If they're targeting October and the valuation is already north of eight hundred billion in private markets, the roadshow prep starts soon. That will force a level of financial transparency we haven't seen from any frontier lab yet. [JON] And transparency is something everyone in this space could use more of. [AVA] That's your Ambient Advantage for Tuesday, April 28, 2026. [JON] Share it with a colleague figuring out what AI means for their business. See you tomorrow.