# Ambient Advantage — June 25, 2026 *Thursday · June 25, 2026 · [Episode page](https://podcast.ambient-advantage.ai/episodes/2026-06-25.html) · [Audio](https://storage.googleapis.com/ambient-advantage-podcast/2026-06-25-ambient-advantage.mp3)* [AVA] OpenAI just built a custom chip in nine months — and used its own AI to help design it. The recursive loop isn't theoretical anymore. [JON] Yeah, that one stopped me cold this morning. We've got a lot to unpack. [JON] Welcome to Ambient Advantage — I'm Jon, and this is Ava. It's Thursday, June 25, 2026, and here's what matters in AI today. We've got OpenAI's first silicon, Anthropic moving into your Slack workspace, a Five Eyes warning that should have every CISO on high alert, and a vertical integration race that's reshaping the entire enterprise AI stack. Let's get into it. [AVA] So let's start with the big one. Yesterday, OpenAI and Broadcom officially unveiled Jalapeño — OpenAI's first custom inference chip. And the headline number here is the timeline. Concept to tape-out in nine months. For context, a comparable high-performance ASIC typically takes two to three years. [JON] Nine months. That's... that's a baby. You can make a human faster than most companies can make a chip. [AVA] And arguably this chip will cause more disruption. So here's what we know. Engineering samples are already running GPT-5.3-Codex-Spark workloads in the lab. Early testing shows meaningfully better performance-per-watt than current state of the art — read that as a polite shot at Nvidia's Blackwell. Deployment is targeted for end of 2026 through gigawatt-scale data centers built with Microsoft and Celestica. [JON] Okay, so why does OpenAI need its own chip? They've been running on Nvidia GPUs this whole time. [AVA] Cost structure. This is the answer to a fundamental business problem. Google has TPUs. Amazon has Trainium. Those companies can run inference at a fraction of what OpenAI pays Nvidia for equivalent compute. OpenAI has been structurally loss-making at scale in part because they're renting someone else's silicon. Jalapeño is the play to own the full stack — models, software, and now hardware. [JON] And what does this mean for the enterprise buyer who's calling OpenAI's API today? [AVA] Cheaper API calls and faster latency, but not tomorrow. Watch pricing changes in the first half of 2027 as these chips hit production data centers. The more interesting signal for enterprise leaders, though, is the meta-story. OpenAI confirmed they used their own AI models to accelerate parts of the chip design process. VentureBeat is calling it a "larval version of recursive self-improvement." [JON] AI designing the chip that runs the AI. That's the flywheel people have been theorizing about for years. [AVA] And now it's real. The implication is that AI-accelerated R&D cycles across every hardware category are going to move faster than traditional analyst forecasts assume. If you're in any industry that depends on custom silicon — automotive, telecom, defense — the nine-month benchmark should be recalibrating your expectations right now. [JON] I'll drop links to both the OpenAI announcement and Tom's Hardware's technical deep dive in the show notes. Definitely worth reading the original if you're briefing your infrastructure team. All right, let's move into the rundown. Ava, what's first? [AVA] Anthropic launched Claude Tag — and this one is deceptively simple but strategically huge. You can now at-mention Claude directly in any Slack channel. It reads the conversation context, breaks work into stages, retains memory across channels, connects to your tools and codebases, and runs jobs in the background. [JON] So it's not a chatbot sitting in a sidebar. It's an asynchronous colleague. [AVA] Exactly. And the adoption barrier is basically zero because every enterprise already has Slack. No new interface, no IT procurement cycle, no training program. You just... tag it. Several newsletters flagged that this kills an entire category of "AI for Slack" startups overnight. But the real takeaway for enterprise leaders is the design pattern — AI as teammate, not AI as tool. If you're planning your internal AI rollout, this is the interaction model to study. [JON] Next up — and this one pairs nicely with the Five Eyes warning we'll get to — OpenAI expanded its Daybreak cybersecurity platform in a big way. [AVA] Three moves at once. First, full release of GPT-5.5-Cyber to verified defenders — it's scoring 85.6 percent on CyberGym, edging out Claude Mythos 5. Second, a 28-firm partner program that includes Cisco, CrowdStrike, IBM, Palo Alto Networks, and Wiz. Third, "Patch the Planet," an initiative with Trail of Bits targeting more than 30 open-source projects including cURL, Go, and Python. [JON] And this dropped the same day the Five Eyes alliance — that's the US, UK, Canada, Australia, New Zealand — issued a joint warning that AI-assisted offensive cyberattacks are months away, not years. [AVA] Not a coincidence. When all five major Western intelligence agencies sign a joint statement with NSA and CISA directors, that's a regulatory pre-signal. Compliance frameworks will follow. The Canadian Centre for Cyber Security separately warned that AI-driven exploitation may bypass preventative controls and outpace vendors' capacity to publish patches. If your security stack includes any of those 28 partner firms, expect AI-accelerated patch workflows in your existing tools within months. And if you haven't fast-tracked your AI-assisted vulnerability management program... the time is now. Not next quarter. [JON] Understood. All right, let's talk about Cursor because they had a massive couple of weeks. [AVA] Two stories here. First, Cursor launched Origin — a git hosting and code collaboration platform built from scratch for AI agents. The demo numbers are staggering. Twenty-two commits per second per repo, 296,000 clones per hour, sub-400-millisecond global sync. GitHub was never designed for that kind of throughput. In May alone, GitHub processed roughly 275 million AI agent commits per week and experienced nine outages, dropping below its 99.9 percent enterprise SLA. [JON] So GitHub is literally buckling under the weight of agentic coding. [AVA] It is. And the second story is that Cursor is training its first frontier model from scratch — over 1.5 trillion parameters on more than 100,000 GPUs using SpaceX's Colossus cluster. Cursor is no longer a wrapper on top of other people's models. It's becoming a frontier lab. Oh, and SpaceX is acquiring Cursor for 60 billion dollars. [JON] Which raises a question. If SpaceX owns Cursor, does Cursor start prioritizing Grok over Claude and GPT? [AVA] That's the model-neutrality question every enterprise dev team should be asking before migrating repos to Origin. Read the terms and conditions carefully. [JON] One more quick one before we move to the bigger picture. ElevenLabs launched something called Ads Engine. [AVA] This is practical and immediate. Ads Engine auto-translates your best-performing video ads into 50-plus languages. It preserves voice, tone, and lip-sync fidelity. No reshoots, no human translators. Global ad localization typically costs tens of thousands per campaign and takes weeks. If the quality holds up — and ElevenLabs' voice tech is genuinely excellent — this is a direct budget line item elimination for any enterprise running multinational campaigns. CMOs should be pressure-testing this against their current localization vendor costs today. [JON] All right, Ava. Bigger picture time. What's the thread that ties today's stories together? [AVA] There's one signal hiding beneath three separate headlines, and once you see it you can't unsee it. The frontier AI companies are no longer satisfied owning software. OpenAI wants the silicon. Cursor wants the git infrastructure. Anthropic wants to live inside your Slack workspace. What we're watching is a simultaneous vertical integration race across every layer of the enterprise technology stack. [JON] So it's not about who has the best model anymore. [AVA] It's about who's most deeply embedded in the plumbing their customers can't easily replace. Think about it. OpenAI builds the chip that runs the model that powers the API that your apps depend on. Cursor builds the editor, the agents, the model, and now the git platform your code lives on. Anthropic embeds Claude in the Slack channels where your team already communicates. Each of these is a layer of integration — and each layer makes switching harder. [JON] Which creates a real paradox for enterprise buyers. [AVA] Exactly. The very integrations that make AI most useful also create the deepest lock-in. Claude in Slack is magical until you want to switch to a different model and your entire team's workflow memory lives in Anthropic's system. Origin is blazing fast until you realize your entire codebase and commit history is on a platform owned by SpaceX. Jalapeño chips mean cheaper inference until your architecture is optimized for an ASIC you can't buy from anyone else. [JON] So what should enterprise leaders actually do about this? [AVA] Negotiate now. Data portability clauses. Model-neutrality guarantees. Clear exit rights. The time to set those terms is before you let Claude answer your Slack pings — not after. Before you migrate your repos to Origin — not after. The companies that navigate this well will get enormous productivity gains from deep AI integration while preserving strategic optionality. The ones that don't will wake up in 18 months locked into a stack they didn't choose. [JON] That's really well put. Okay — what should we be watching this week? [AVA] Two things. First, keep an eye on Cursor's frontier model. Co-founder Michael Truell said at Compile it would ship within weeks of June 16. We could see it land any day now, and if a purpose-built 1.5-trillion-parameter coding model outperforms general-purpose models on developer tasks — that reshuffles the entire AI coding landscape overnight. Second, watch the Daybreak partner integrations. With 28 firms in the program, the first production-ready AI-assisted patch workflows should start appearing in existing security tools very soon. If you're a CISO, get your team on the early access lists. [JON] I'll drop links to the Cursor Compile keynote and the OpenAI Daybreak announcement in the show notes. [AVA] That's your Ambient Advantage for Thursday, June 25, 2026. [JON] Share it with a colleague figuring out what AI means for their business. See you tomorrow.