🌟 Editor's Note: Recapping the AI landscape from 06/02/26 - 06/29/26.
🎇✅ Welcoming Thoughts
Welcome to the 1st edition of KPAI Monthly.
What’s included: company moves, a weekly winner, AI industry impacts, practical use cases, and more.
What an exciting time in AI, good to be back!
Next issue will mean the newsletter is 1 year old and we will crown the winner of the initial AI Race.
I’ve been using Wispr flow a bunch, it’s so easy to just talk to my terminal like a person and have it know what to do.
We’re starting to move to a heavyweight battle between OpenAI and Anthropic, the two best models right now by a long-shot.
The AI models are starting to become less important than the framework and harnesses around them.
Used Gemini today for the first time in a while, it’s pretty bad.
GLM 5.2 is apparently very impressive (huge open-source model).
In addition to leading open-source models, users in China are reselling tokens at extreme discounts. Probably best to stay away.
98% of OpenAI employees are now working out of Codex. (Same here, but Claude Code).
Pretty unimpressed with Google right now, they’ve fallen off a bit.
Anthropic and OpenAI are absolutely killing it.
AI Race section got a bit long, I’ll shorten it next month.
I like the Monthly newsletter cadence.
Let’s get started—plenty to cover this Month.
👑 This Week’s Winner: OpenAI // ChatGPT
OpenAI takes the month. Between a frontier model launch, a confidential IPO filing, and its first custom chip, OpenAI had an incredibly strong June. Here's the recap:
GPT-5.6 Previewed: OpenAI started a limited preview of GPT-5.6 in three tiers: Sol (flagship), Terra (balanced), and Luna (fast, low-cost). Sol is its most capable model yet, adding a "max" reasoning effort and an "ultra" mode that runs subagents on long tasks. Up there with Claude Fable 5 as the most capable model ever built, but restricted for public use by Uncle Sam.
Filed to Go Public: OpenAI confidentially filed its IPO paperwork with the SEC at an $852 billion valuation, with Goldman Sachs and Morgan Stanley leading the process, just a week after Anthropic did the same. A fall listing is on the table, though OpenAI says the timing is still undecided. Will be an exciting IPO along with Anthropic’s. Very impressed with OpenAI’s 2026 after a lackluster 2nd half of 2025.
First Custom Chip, "Jalapeño": OpenAI and Broadcom unveiled Jalapeño, OpenAI's first chip built specifically for running AI models, taken from design to production in about nine months. It is the first piece of a multi-generation effort to lean less on NVIDIA, with initial deployment targeted for late 2026. Still won’t have any meaningful impact on NVIDIA market share.
OpenAI also brought Codex Remote to general availability for all paid plans, letting people drive long coding sessions from their phone, and shipped a physician-vetted health upgrade to ChatGPT alongside a new Partner Network aiming to certify 300,000 consultants. The new model also drew Washington's attention: at the administration's request, OpenAI limited GPT-5.6 to about 20 government-approved companies.
From Top to Bottom: Open AI, Google Gemini, xAI, Meta AI, Anthropic, NVIDIA.
⬇️ The Rest of the Field
Who’s moving, who’s stalling, and who’s climbing: Ordered by production this week.
🔴 SpaceX // Grok
Largest IPO in History: SpaceX, which now owns xAI, completed the biggest IPO ever on June 12, raising about $75 billion at a $1.77 trillion valuation. Shares jumped on debut, peaked at $225 on June 16, then fell about 32% by late June. As long as Elon is around this company is going to be incredibly successful.
Buys Cursor, Reloads Grok: SpaceX acquired the AI coding tool Cursor for about $60 billion and moved a few dozen of its top Starship and Starlink engineers onto Grok. The first result is Grok 4.5, a 1.5-trillion-parameter model now in beta at Tesla and SpaceX. Love this move. SpaceX may one day compete with Claude Code and Codex.
Rivals Renting Colossus: SpaceX is now leasing its Colossus compute to three competitors: Anthropic at $1.25 billion a month, Google at $920 million, and Reflection AI at $150 million. Combined, that is more than double SpaceX's entire 2025 revenue. Smart play.
🟠 Anthropic // Claude
Launched Fable 5, Forced Offline: Anthropic released Claude Fable 5, its most powerful public model, on June 9, and three days the US Govt. forced it and Mythos 5 offline. The White House cleared Mythos 5 for 100-plus vetted US organizations on June 26; Fable 5 is still under review. Very impressive model, we had a good 3 day run! Should be back soonish.
Claude Tag in Slack: Anthropic launched Claude Tag, a way to drop Claude into a Slack channel as a shared teammate that takes a task, breaks it into steps, and works through it on its own. Cool concept, another notable shift to where AI is being communicated with.
Seoul Office and Korea Push: Anthropic opened a Seoul office, its third in Asia, and lined up Claude deployments with NAVER, Samsung SDS, and LG CNS, plus an agreement with Korea's science ministry on AI safety.
⚪️ NVIDIA
The Age of Agents: At its annual shareholder meeting, Jensen Huang declared an "age of agents" and said international revenue topped $30 billion across about 40 countries. He called black-market data centers built from smuggled chips a "dead end." Don’t agree with the last statement but it certainly is not meaningful with regard to scale.
Vera Rubin in Full Production: NVIDIA's next platform, Vera Rubin, reached full production, pairing the Rubin GPU with Vera, which it calls the first CPU built for AI agents, at roughly 10x the throughput of the prior generation. OpenAI, Anthropic, Meta, and Microsoft are named adopters. NVIDIA still doing NVIDIA things.
Supercomputers in Europe: NVIDIA said a record 35 of its AI supercomputers are being built across 23 European countries, the region's largest single-year buildout, serving more than 3 million researchers. And more to come.
🟣 Google // Gemini
Gemini 3.5 Flash Goes Default: Google made Gemini 3.5 Flash the default across Search AI Mode, the Gemini app, and its API. Also added computer use. AI Mode passed 1 billion monthly users, a reach no rival has. Lots of usage, but a quiet month. No new meaningful model in quite some time. Unimpressed.
Gemini → Antigravity CLI: Google retired its Gemini command-line tool and moved developers to the new Antigravity CLI, part of its agent-first developer platform, keeping features like skills, hooks, and subagents. Antigravity still a ways behind Codex and Claude Code.
Caps Meta's Gemini Access: Google limited how much Meta could use its Gemini models after Meta asked for more capacity than Google could supply, a rare public sign of Google's own compute crunch. Compute is still gold.
🔵 Meta // Meta AI
OpenSource Mind Reader: Meta's research lab showed Brain2Qwerty v2, a non-invasive system that reads brain activity through a helmet-style scanner, with no implant, and turns the sentences someone is trying to type into text at 61% accuracy, up from about 8% for earlier methods. Dystopic, awesome.
Meta Predictions: Internal documents show Meta is building a standalone app called Arena where people bet play money on real-world events, with its Llama model generating and settling the markets. Meta has had no luck building frontier AI models so they’re just finding things to do.
Morale Near a 20-Year Low: CTO Andrew Bosworth told staff morale is close to the worst he has seen in two decades after Meta cut about 8,000 jobs and reshuffled engineers into a new Applied AI unit, writing that leadership did "an atrocious job" explaining the reorganization. Zuck seems a bit lost right now.
🤖 Impact Industries 💻
Robotics // Amazon Proteus
Amazon unveiled a new version of its Proteus warehouse robot that takes plain-language instructions. A worker tells it what needs doing, and the robot figures out the priority, the route, and the timing on its own, moving carts that weigh up to 400 kilograms around the floor. The current Proteus already runs in 25 US fulfillment centers, and this conversational version is being tested now, with a European rollout planned for the first half of 2027.
Read the Story
Dev Tools // Open Models
Open AI models are getting good enough to rival the big paid ones, and you can download and run them on your own computer. The clearest example is GLM-5.2, a free model from Chinese lab Z.ai that now ranks as the top open model and beats some paid ones on coding. Models like this run locally, even on something the size of a Mac mini, which worries the major AI companies: their business is selling access, and a model on someone's own machine is hard to regulate.
Read the Story
💻 Interview Highlight: Jensen Huang with Sequoia Capital
Interview Outline: Jensen Huang details a fundamental architectural shift in global computing, moving from 60 years of file retrieval to real-time intelligence generation. He explains how artificial intelligence has evolved from an interesting novelty into autonomous, agentic systems capable of performing useful work, effectively functioning as digital employees earning an hourly wage. Huang outlines the physical layout of "AI Factories," emphasizes the massive ROI of the underlying infrastructure, and frames the multi-trillion-dollar scale of the opening application layer.
About the Interviewee: Jensen Huang is the co-founder, President, and CEO of NVIDIA. A visionary pioneer in accelerated computing, his early strategic pivot to GPUs and the proprietary CUDA platform positioned NVIDIA at the absolute epicenter of the global artificial intelligence boom, driving the company to become one of the most valuable enterprises in human history.
Interesting Quote: "One rack weighs 2 tons, costs $4 million, and contains 1.5 million parts... It's the most expensive piece of equipment in the world, but we mass-produce them like we manufacture mobile phones."
Condensed Interview Highlight — Jensen Huang (Sequoia Capital AI Briefing)
1. Konstantine Buhler: How has AI evolved from a novelty to a business necessity?
Jensen Huang: Two years ago, ChatGPT was an amusing proof of concept, but today we have fully agentic systems that can reason, break down problems, and do useful digital labor. We are moving into a model where enterprises are essentially hiring AI by the hour, paying it $20 to $30 for the specific tasks it executes. The rapid adoption we're seeing across the board is driven by the fact that AI now produces tangible, measurable business outputs rather than just answering basic prompts.
2. Konstantine Buhler: How is the core function of computing changing?
Jensen Huang: We are undergoing the single largest architectural shift in computing in 60 years, moving away from file retrieval to real-time generation. In the old model, data centers were essentially storage units built to hold pre-existing files that computers would simply look up and retrieve whenever a user made a request. Now, every word, line of code, image, and video is manufactured dynamically from scratch and fully customized for whoever is asking.
3. Konstantine Buhler: Should people be worried about AI replacing their jobs?
Jensen Huang: History shows that automation does not lead to mass unemployment; instead, it comprehensively increases total labor demand and elevates existing professions. When tasks become automated, companies naturally expand their operations, which creates a need for higher-level strategic work. You will not lose your job to artificial intelligence itself, but you very well might lose it to a human professional who knows how to leverage AI effectively.
4. Konstantine Buhler: How should companies structure their investments in this space?
Jensen Huang: The entire modern artificial intelligence industry can be visualized as a highly interconnected, five-layer cake. It starts at the foundational level with clean energy, moving up through specialized silicon chips, physical factory infrastructure, foundation models, and finally, consumer-facing applications. Each individual layer represents a distinct multi-trillion-dollar ecosystem filled with its own unique partners, engineering standards, and technical workflows.
5. Konstantine Buhler: What is the next frontier beyond text and software?
Jensen Huang: Artificial intelligence can learn the fundamental structure of anything, provided we have the data to model it, including the core laws of physics. Just as it learned to understand human language, the technology can now learn the underlying meaning of proteins, the significance of genes, and the actual functions of living cells. This pivot is taking us far beyond basic software chatbots and moving us directly into physical robotics and advanced biology.
👨💻 Practical Use Case: Agent Harnesses
Difficulty: Advanced
You have probably noticed that the same AI model can feel brilliant in one app and useless in another. More often than not, that comes down to the harness around it rather than the model itself. On its own, a model only generates text. An agentic harness is the software that wraps that model and lets it actually get things done: take actions, use your tools, and work through a task one step at a time instead of answering once and stopping. A few pieces make that possible:
With Projects, you can:
The loop: The model suggests a step, the harness carries it out, then hands the result back so the model can decide what to do next. That cycle repeats until the job is finished.
Execution: The actions the model is allowed to take, such as reading a file, running code, searching the web, or calling another service. The harness runs each one and returns what it finds.
Context: What the model can keep in view at any one moment. Because that space is limited, the harness decides what to feed in and what to set aside as the work grows.
Governance: The permissions and limits around all of it, including a human sign-off before anything risky. This is what keeps an agent from quietly doing damage.
This is why the harness matters as much as the model underneath. Claude Code, Codex, and Cursor all run on the same handful of models, yet they feel very different to use, because each one wraps those models in a different harness.
🩻 Startup Spotlight
Midjourney Medical
Midjourney Medical — Full-body imaging without the MRI price tag.
The Problem: Traditional MRIs and CT scans are expensive, slow, and trapped inside clinical bottlenecks. Getting a clear view inside your body usually requires either a massive medical event or thousands of dollars out of pocket. Plus, CT scans expose you to radiation, and MRIs require powerful, high-risk magnetic fields.
The Solution: A 60-second full-body scanner called "Ultrasonic CT." Despite the name, it involves zero radiation or magnets. Instead, you step onto a circular platform and descend into a shallow pool of water ringed with half a million ultrasound sensors. The system fires sound waves through your body from every angle, using two petaflops of processing power to reconstruct detailed 3D maps of your muscle, bone, fat, and organs in real time.
The Backstory: In a massive pivot from text-to-image AI, Midjourney CEO David Holz launched this healthcare division on June 17, 2026. Rather than inventing the physics from scratch, Midjourney signed a $74 million co-development deal with Butterfly Network to use their advanced ultrasound-on-chip technology. The long-term play? A fleet of 50,000 scanners worldwide by 2031. To bypass initial regulatory hurdles, they are launching the tool at a flagship "Midjourney Spa" in San Francisco in late 2027.
My Thoughts: This is one of the most exciting releases since ChatGPT. I think one of the things future generations will comment about ours is how primitive our health tech was. This includes both prevention, as well as treatment. I believe AI enabled research will lead to tons of opportunity and innovation in this space, and Midjourney, an AI image company, just offered up a starting point.
“It’s not likely you’ll lose a job to AI. You’re going to lose the job to somebody who uses AI”
- Jensen Huang | NVIDIA CEO
What’s your current model of choice? Till Next Time,
Noah from KPAI