OpenAI Snags Noam Shazeer and Builds a Massive Partner Network

Gemini Co-Lead Leaves Google for OpenAI

Noam Shazeer, VP of Engineering at Google and one of the two minds behind Gemini, announced on X on June 18, 2026, that he is leaving the company to join OpenAI. He did not specify his new role, but the move is significant: Shazeer co-led Gemini and has been in Google's AI engine room as VP of Engineering since August 2024. His resume is that of one of the most important figures in the field: a software engineer at Google from 2000 to 2021, then CEO of Character.ai, returning in August 2024 through a licensing deal between Google and Character.ai. Google's response was brief, stating they were “grateful for Noam's significant contributions over the years.” With this, the Gemini team loses one of its two captains to a direct competitor in the middle of a critical phase. → AI Secret

Synthszr Take: Shazeer isn't just anyone; he is a co-inventor of the Transformer architecture, the 'T' in GPT. The fact that this particular person is moving from Google to OpenAI has an almost poetic quality: the blueprint that Google published in 2017 and then left behind is now arriving at its competitor via its own architect. In April, we wrote that OpenAI wanted to become the new Apple, and for that, it needs not just data centers but the few dozen minds who know how to build the next generation. The real moat in this industry is neither patents nor GPUs, but a handful of people you can't simply order more of. Google paid a high price to bring Shazeer back in 2024 and lost him again 22 months later—that's an expensive revolving door. Anyone who wants to stay ahead in the AI race should invest less in marketing budgets and more in understanding why the best people stay or leave.

OpenAI is Building a Consulting Army of 300,000

OpenAI is launching the Partner Network, a global program for system integrators, strategy consultancies, and tech and data firms to build, sell, and deploy solutions around its models. $150 million is being invested in the ecosystem, and by the end of 2026, OpenAI aims to train 300,000 certified consultants. The program has three tiers (Select, Advanced, Elite) with high hurdles for sales performance, technical expertise, and deployment experience. There are also specializations in Codex, Cybersecurity, and Agents, as well as a pilot called Forward Deployed Experts, which connects partners directly to OpenAI's own Forward-Deployed-Engineering teams. OpenAI's reasoning: the company's bottleneck is no longer the models, but identifying the right use cases, redesigning workflows, and change management. No single company can deliver every solution in every market, hence the ecosystem approach. → AI Secret

Synthszr Take: OpenAI has understood that the margin isn't in the model, but in the last mile to the customer. This is the exact same play we described in early May when OpenAI followed the Palantir pattern with Forward Deployed Capital. Now comes the industrialized version: 300,000 certified consultants form a sales machine that turns Accenture, Deloitte, and McKinsey into resellers without OpenAI having to pay for the consulting hours itself. This is the real moat, not GPT-5.3-Codex, because models are commoditized on a quarterly basis, while customer relationships and trained practitioners are not. Any German consulting or digital agency wondering if the effort for an Elite certification is worth it should realize that the lock-in works both ways: OpenAI locks in the consultants, and the consultants lock their clients into OpenAI. The more interesting question is what Anthropic and Google will do now, as a once-established partner ecosystem cannot be copied overnight. Whoever occupies the implementation playing field early wins the enterprise budgets for the next five years.

Midjourney is Now Building Body Scanners

Midjourney, previously known for text-to-image, has announced its own division for medical imaging. You step into a pool of water, sound waves pass through your body from all angles, and after 60 seconds, a 3D map of your entire body is available: no radiation, no magnetic fields, just ultrasound. The specs are extraordinary: 500,000 transducers fire simultaneously, 17 gigabytes of data are processed per second, and a scan costs a few dollars instead of several hundred. The first location is set to open in a San Francisco spa in 2027, initially providing body composition data while they work towards FDA approval for full diagnostics. In parallel, Anthropic has beefed up Claude Design: it imports your actual design system from a GitHub repo and checks its own output against your components before you see it. The pattern is clear: AI labs are completely swallowing adjacent industries instead of just embellishing them with features. → AlphaSignal

Synthszr Take: An image generator that, without ever having built a physical product, suddenly promises a full-body scanner for a few dollars sounds like a PowerPoint illusion. But the real story is the speed. The research depth of these labs is accelerating its translation into products, and they are gravitating towards markets previously held by a few large imaging corporations with hardware moats. When a software house shortens the path from idea to industrialization this drastically, every established manufacturer that still thinks in terms of years instead of quarters comes under pressure. The 2027 deadline is the true litmus test, as atoms are less forgiving than pixels (FDA approval and hardware scaling can't be provisionally patched in a sprint). Anyone watching this and waiting for the next strategy offsite has not understood the speed of the competition. The question is not whether these bets will pay off, but who outside the AI world can even keep up.

Slack Aims to Become the Control Layer for Agents

Slack has released an IDC white paper by analyst Arnal Dayaratna that sells a simple thesis: AI agents are only as good as the context they have. According to the survey, 52% of developer organizations have already adopted agents, with 66% of those in production. Slack's argument: its architecture bundles conversation history, workflow activity, integration signals, and rights-aware access in the same place where teams already communicate and make decisions. Agents without visibility into business rules, past decisions, and customer context create risk instead of value. Slack is explicitly positioning itself as the coordination layer for multi-agent work, including human oversight. The paper provides the sales arguments right along with it: less coordination effort for knowledge workers and developers. → Techpresso

Synthszr Take: Slack has understood where the real money is, and it's not in the next copilot window. Whoever owns the control plane—the layer that brings together orchestration, permissions, and human approval—has the moat. That's exactly what they're claiming here: not an agent that sits alongside work, but the environment in which the agent fleet runs and humans sign off. This fits with what we noted in May when AI labs started building agents instead of just models. However, the 66% production rate in the paper should be taken with a grain of salt; it's Slack's own survey, not a neutral market report. It will be interesting to see if Slack can truly hold this position or if Notion, Microsoft, and native agent platforms will claim the coordination layer themselves. Whoever loses there will end up just selling an expensive chat tool.

Google and Microsoft Team Up Against Anthropic and OpenAI

A new backend protocol might sound like the most boring topic an IT department deals with. Yet, it's worth a look: Google, Microsoft, Salesforce, Snowflake, ServiceNow, along with Hugging Face and Databricks, agreed on Wednesday to a standard called Agentic Resource Discovery (ARD). The idea: an employee asks GitHub Copilot, Gemini, or Salesforce, and the app independently finds any AI function the company is already paying for, provided the respective provider participates in ARD. It's an extension of the Model Context Protocol that Anthropic invented last year. Anthropic and OpenAI are not on the list of initial supporters because they want to establish Claude and ChatGPT as the central hub in the employee's mind. Microsoft promises that ARD will make GitHub Copilot significantly cheaper because the tool will waste less computing power finding suitable apps. In parallel, Cursor, with its planned code-hosting product Origin, confirms exactly what Microsoft had warned about: GitHub's core business as a repository provider is on shaky ground. → Applied AI

Synthszr Take: Whoever owns the access, owns the business, and that's exactly what this is about. The established providers have only one real moat: their installed base of millions of Office, Salesforce, and ServiceNow seats, and ARD is the attempt to save this moat in the AI era before Claude and ChatGPT become the first app of the workday. The cost argument is the real leverage: cost-weary IT folks don't jump on visions, they jump when Copilot burns less compute. What's interesting is that the same Microsoft that co-sponsors ARD is, according to our note from mid-June, considering DeepSeek for Copilot and switching to usage-based pricing: compute discipline is the theme everywhere. Whether ARD succeeds depends not on the protocol, but on whether the insurgents play along, and they have no reason to. Meanwhile, the Cursor-Origin situation shows how quickly a lock-in can crumble when your own product constantly fails and a competitor makes switching easy. In the end, standards are won by whoever delivers the cheapest and most stable experience, and that's exactly where GitHub has a problem right now.

Vercel Introduces Eve, an Open-Source Agent Network

Vercel has introduced Eve, an open-source framework for building, running, and scaling agents. The provider, previously known mainly for its frontend deployment and the Next.js ecosystem, is thus pushing deeper into the agent stack. Eve covers the entire lifecycle: from initial prototyping to production operation and scaling. The fact that it comes as open source is the real statement, as Vercel is giving away the framework and earning from the underlying infrastructure. This places Vercel in a trend we've been observing for months: from Cloudflare's agent tooling to the labs that have long been building agents instead of just models. The market for agent frameworks is filling up quickly. → The Deep View

Synthszr Take: Vercel is doing exactly the right thing here, and it's in its own self-interest. The framework is free, but the data centers underneath it are not. Anyone using Eve builds their agents on a logic that almost inevitably runs on Vercel's infrastructure, and that's where customer loyalty is created. In May, we wrote that agents are becoming more important than models. Now we're seeing the next level: the layer that orchestrates, operates, and scales agents is becoming the real moat question. It's interesting that a frontend company is occupying this position and not one of the big labs. Anyone seriously building agents today should try Eve in the next two weeks before the stack decision for the next three years is made.

Databricks Bets on Open Ecosystems to Combat Agent Chaos

Almost all major tech companies are currently racing in the same direction: bringing the AI agent into the enterprise. The Deep View has documented the attempts by Microsoft, Snowflake, AWS, Anthropic, OpenAI, and Perplexity, and the same problems appear everywhere: data leakage, compliance gaps, data privacy, security vulnerabilities, inaccuracy, and soaring token costs. The real problem, however, is that the agents are not yet delivering the promised value that would justify these costs. On top of that, there's a new annoyance: every software service is pushing its own agent, its own builder, its own MCP server, creating customer lock-in without compatibility. At the Data + AI Summit on Tuesday, Databricks CEO Ali Ghodsi called it simply a complete mess. Databricks' answer is, on one hand, its own agent called Genie (with Genie One, Ontology, App Builder, Agent, and Code), and on the other hand, the principle of open ecosystems. Chief AI Scientist Jonathan Frankle puts it bluntly: customers can pull their data out of Databricks completely at any time because there's hardly any lock-in, and that's precisely the plan. → The Deep View

Synthszr Take: The interesting move here isn't Genie, but the promise of openness surrounding it. Ghodsi is selling openness as a feature while everyone else is building a walled garden, and in a world full of proprietary agent silos, that is the moat. Frankle's statement that he has to earn the business anew every day is the most honest sales strategy in a long time (and also the smartest, because it takes the immune system of large enterprises with their legacy systems seriously). But beware of too much enthusiasm: the real problem with agents isn't the tool layer, but the question of who is allowed to act on an agent's response and who is liable when it's wrong. Gartner expects about 30 percent of generative AI projects to die after the proof of concept, almost never because of the technology, but almost always due to a lack of decision-making authority. Open data portability doesn't solve the interpretation problem; it just makes it cheaper to switch providers when you realize the bottleneck is elsewhere. If you want to make agents productive, you should build the governance first and then choose the tool, and for the latter, freedom from lock-in is a damn good selection criterion.

Microsoft is Selling OpenAI Models to China and Making a Fortune

Microsoft has built a substantial business selling AI models to Chinese corporations amidst the growing rivalry between the US and China. According to Bloomberg, the biggest customer is none other than ByteDance, the Beijing-based parent company of TikTok, which predominantly uses OpenAI models. The Beijing company is on track to spend more than a billion dollars a year on Microsoft's AI and cloud services. All of this is happening while Washington and Beijing are taking an increasingly hard line on artificial intelligence, with export controls setting the tone. The people who confirmed this wished to remain anonymous because it is a confidential matter. ByteDance obtains the models through Microsoft's Azure infrastructure, not directly from OpenAI. → Marcus Schuler

Synthszr Take: While politicians in Washington talk about decoupling, Microsoft is collecting a billion dollars a year from the parent company of TikTok. This is the moat of the cloud business in its purest form: the models are interchangeable, but the infrastructure and the contractual relationship are not. In April, I wrote that Microsoft was undermining its partner OpenAI with its own models. Now we see the other side of the same coin, as those same OpenAI models are financing Microsoft's growth in China, a market OpenAI itself cannot access directly. The interesting question is how long Redmond can maintain this balancing act before a new round of export controls cuts off the pipeline. As long as compute is scarce and margins are fat, Microsoft will quietly book every geopolitical contradiction. Whoever owns the infrastructure sells to both sides, and the models are just the powder flowing through the pipe.

Bernie Sanders Wants to Redistribute $7 Trillion from the AI Industry to the People

Bernie Sanders has introduced a bill that would create a sovereign wealth fund modeled after Norway's, financed by a one-time 50 percent tax on the stocks of the largest AI companies. Any company with an annual AI revenue of $200 million or more would be affected, with new companies automatically included upon reaching this threshold. Sanders estimates the fund's volume at $7 trillion, from which every American would receive over $1,000 annually as a 5 percent dividend, plus funds for health, education, and housing. A seven-member commission, confirmed by the Senate, would manage the fund through voting shares and could block corporate decisions that harm the public. In a conversation with Sanders, Sam Altman's views were far from his, and Trump's former AI advisor David Sacks called the bill “outright expropriation.” Particularly tricky is the requirement to separate non-AI business from AI business, which would affect Elon Musk's convoluted structure of xAI, X, SpaceX, and possibly Tesla. In the Republican-dominated Congress, the bill is considered to have no chance without Trump's approval. → Ars Technica

Synthszr Take: Sanders himself knows this bill won't pass Congress, and that's precisely why it's interesting. It shifts the Overton window. In June, Trump himself toyed with the idea of taking state shares in AI companies, and suddenly Sanders' 50-percent hammer sounds like the radical end of a debate that wouldn't even be happening without him. The interesting number isn't the $7 trillion estimate (that's just a guess), but the 5 percent that Altman would voluntarily give back to the people. Sanders aptly calls this a buyout, and he's right: anyone who disrupts millions of jobs while distributing a pittance to society is buying quiet, not legitimacy. For Europe, there's a lesson in this that has nothing to do with nationalization: if you remain dependent on others for compute, capital, and your own models, you won't even have the 5 percent to negotiate at the end of the day. Building your own capabilities beats any redistribution plan, because you can only share what you own.

Unreal Engine 5.8 Wires Language Models Directly into the Game Engine

Epic Games has shipped Unreal Engine 5.8, the last planned major release of the UE5 line. Included is an experimental plugin based on the Model Context Protocol (MCP), which connects any LLM directly to the engine's core systems, such as Blueprints, Assets, Levels, Materials, and Meshes. In a demo video, Epic shows how Anthropic's Claude Code uses the plugin to pull objects from an asset library, arrange them in a scene, and adjust the lighting to match real reference images. Other features include Mesh Terrain for more complex worlds, MegaLights with a 60fps target on current consoles, and MetaHuman Animator with full-body capture from a single camera. Epic frames the plugin as groundwork for Unreal Engine 6, where LLM integration will become a central part of the creation pipeline. The stated goal for UE6 is to reduce tedious content authoring and allow more time for creative iteration. Early Access is targeted for late 2027, with the full launch 12 to 18 months later. → Techpresso

Synthszr Take: MCP is eating its way through the entire toolchain, and faster than most expected. In February, WebMCP for browser agents was the headline; now the same protocol is inside the engine used to develop a significant portion of all AAA titles. What Claude Code demonstrated for codebases (read, write, execute, correct in loops) is now happening with the 3D scene: the language model becomes part of the workflow, not a chat window next to it. The timeline is interesting. Epic is laying the foundation with 5.8 and pushing serious integration to UE6 with Early Access in late 2027, which is an eternity at this speed. Anyone building asset pipelines or level designs should get their hands on the MCP plugin now and align their own toolbox with it, instead of waiting for 2027. The engine is becoming an aggregator for any connectable LLM, and the moat is shifting from rendering to the question of who can orchestrate the creation loop most naturally.