The battle for AI infrastructure is heating up
June 13, 2025
Good morning. Yesterday's massive internet outage proved that Google, YouTube, Spotify and half the web are basically one big house of cards—and Google Cloud's identity system was the card that fell. Nothing says "robust digital infrastructure" quite like the entire internet having a synchronized meltdown.
— The Deep View Crew
In today’s newsletter:
🧠 AI for Good: This brain implant gave an ALS patient his voice back
💰 OpenAI employees have cashed out $3 billion in company shares
☁️ The battle for AI's infrastructure is heating up, and it's not just about chips
🧠 AI for Good: This brain implant gave an ALS patient his voice back
Source: UC Davis Health
Casey Harrell lost his ability to speak to his four-year-old daughter as ALS gradually paralyzed him. For two years, she couldn't understand her father's words. Then, everything changed.
What happened: Researchers at UC Davis Health successfully restored Harrell's voice using a brain-computer interface — a system that reads brain signals and translates them into computer commands — combined with AI that converts neural activity into speech with up to 97% accuracy, the most accurate system of its kind.
Neurosurgeon David Brandman implanted four microelectrode arrays — tiny grids of sensors just 3 millimeters wide — into the left precentral gyrus, the brain region that controls speech muscle movements. The arrays record electrical activity from 256 individual electrodes as Harrell attempts to speak.
How the AI works: The system uses multiple layers of AI to decode speech. First, machine learning algorithms analyze electrical signals from 256 tiny electrodes implanted in Harrell's brain. "We're really detecting their attempt to move their muscles and talk," explained neuroscientist Sergey Stavisky, co-director of the UC Davis Neuroprosthetics Lab.
The AI then translates these brain patterns into phonemes — the basic units of speech sounds like "ba" or "ka" — before assembling them into complete words. The breakthrough uses BlackRock Neurotech's NeuroPort system combined with the publicly available Backend for Realtime Asynchronous Neural Decoding platform to process signals in real-time.
Most importantly, a separate AI model trained on old recordings of Harrell's voice from before his diagnosis recreates his natural speaking voice, not a robotic substitute.
At the first session, "he cried with joy as the words he was trying to say correctly appeared on-screen. We all did," Stavisky said.
The breakthrough: The AI required minimal training to achieve remarkable results. Within 30 minutes of the first session, the system achieved 99.6% accuracy with a 50-word vocabulary. After just 1.4 hours of additional training data, the AI expanded its vocabulary to 125,000 words with 90.2% accuracy — learning at a pace that would be impossible for traditional speech recognition systems.
Why it matters: Harrell now uses the system up to 12 hours daily to speak with family, participate in Zoom meetings and work full-time. The research, published in the New England Journal of Medicine, offers hope for millions with ALS, stroke and other conditions that rob people of speech.
"Something like this technology will help people back into life and society," Harrell said.
The Remote Agent for Professional Devs
You ship to production; don’t trust your code to luck (or to just any background agent).
Augment Code's powerful AI coding agent meets professional software developers exactly where they are, delivering deep context and production-ready capabilities. And now Augment Code Remote Agents are available in VS Code to help your team clear out its backlog.
Remote Agents can tackle technical debt, refactor code, and update your tests and docs, all from the cloud while you and your team build what’s next.
Sign up and let Augment Code transform how your team works (no luck required)
💰 OpenAI employees have cashed out $3 billion in company shares
Source: ChatGPT 4o
OpenAI employees and former staff have sold nearly $3 billion worth of company shares since 2021, an unusually large amount for a six-year-old startup that has never gone public.
What's happening: SoftBank bought about $240 million in shares from a small group of current and former OpenAI employees this spring, according to The Information. This brings SoftBank's total purchases of employee shares to around $1.7 billion since January.
The sales happen through tender offers — organized events where the company allows employees to sell their shares to outside investors. OpenAI has been holding these sales roughly twice a year, with employees able to sell anywhere from $2 million to $10 million worth of shares each time.
Why this matters: Since OpenAI isn't publicly traded, employees can't simply sell their stock on the open market like they could at Apple or Microsoft. These organized sales give workers a way to cash out equity compensation without waiting for an initial public offering that may never come.
The participation rates are striking. In August 2021, 90% of eligible employees sold shares at $52 each. By January 2025, about 74% participated at nearly $210 per share.
OpenAI is currently raising $40 billion at a $260 billion valuation, with shares priced at more than $250 each. Employees haven't yet been able to sell at this new, higher price but likely expect another tender offer soon.
The What, Why, and How of MCP
If you aren’t an expert on MCP (Model Context Protocol), you’re not alone. We all have questions about this buzzy new system, and no one is more equipped to answer them than two seasoned experts in software engineering – which means you’re in luck.
Join Prismatic’s Chief Software Architect, Justin Hipple, and Manager of Solutions Engineering, Jake Hagle, as they break down what MCP can do, how it can power your AI integrations, and what it looks like in action.
Meta has finalized a $14.3 billion purchase of a 49% stake in data labeling startup Scale AI
Barbie-maker Mattel and OpenAI are teaming up
AI chatbots are still hungry for books—and libraries are stepping up
Spain’s Multiverse just raised €189 million (around $217 million) to shrink AI models faster
Prediction market Kalshi aired a 100% AI generated ad in Game 3 of the NBA finals
Could AI really be trusted with war–and–peace decisions?
Coco Robotics, backed by Sam Altman, just raised $80 million
ByteDance just released Seedance 1.0 a new impressive video gen model
The right hires make the difference. Scale your AI capabilities with vetted engineers, designers, and builders—delivered with enterprise rigor.
AI-powered candidate matching + human vetting.
Deep talent pools across LatAm, Africa, SEA.
Zero upfront fees—pay only when you hire.
☁️ The battle for AI's infrastructure is heating up, and it's not just about chips
Source: ChatGPT 4o
The AI infrastructure market just got a lot more interesting. While AMD mounted its most serious challenge yet to Nvidia's chip dominance this week, Nvidia simultaneously doubled down on a completely different strategy: controlling the entire platform layer of AI development.
At a launch event in San Jose, AMD CEO Lisa Su unveiled the company's next-generation Instinct MI400 AI chips with a surprise guest. OpenAI CEO Sam Altman appeared on stage to announce his company will use the new processors. "It's gonna be an amazing thing," Altman said.
This is a big deal. OpenAI has been one of Nvidia's most prominent customers, and Altman's endorsement gives AMD serious credibility in its fight against Nvidia's 90% market dominance.
But here's where it gets complicated. Even as AMD attacks Nvidia's hardware business, Nvidia is building something potentially more durable: a platform that makes it the central hub for AI development, regardless of whose chips developers actually use.
AMD's hardware challenge is real
AMD's new MI400 chips represent a genuine technological leap. Unlike regular computer processors that handle tasks one at a time, AI chips contain thousands of smaller processing cores designed for the massive parallel computations that power artificial intelligence. They excel at matrix multiplication — the fundamental math behind neural networks.
The MI400 chips will be assembled into complete server racks called Helios, enabling thousands of chips to work together as a unified system. AMD's approach combines its graphics processing units with its central processing units and networking chips from its 2022 Pensando acquisition.
The customer wins are impressive:
Oracle plans to deploy clusters with over 131,000 AMD chips
Meta uses AMD processors to run inference for its AI models
Microsoft relies on them for Copilot features
AMD's pitch is simple: "significant double-digit percentage savings" over Nvidia chips through lower power consumption and aggressive pricing.
Nvidia's platform play
While AMD focuses on building better chips, Nvidia is executing a different strategy entirely. The company's new DGX Cloud Lepton marketplace gives developers direct access to server-grade chips through a centralized Nvidia interface.
Here's how it works:
Aggregates GPU inventory from providers like CoreWeave, Nebius, and Lambda Labs
Customers use a single Nvidia account to rent and manage compute from multiple providers
Nvidia bundles software tools for monitoring, training, and scaling AI workloads
Uses infrastructure from Nvidia's Brev.dev acquisition to make cloud setup easier
This removes friction from the supply chain while positioning Nvidia as the central hub for AI development. Not everyone is happy about this. Several GPU providers chose to join the platform to stay competitive, but others, like Parasail, declined due to concerns about losing customer relationships and pricing control.
Nvidia's confidence amid the challenges
Nvidia CEO Jensen Huang seems remarkably unconcerned about the competition. Speaking to analysts this week in Paris, Huang said he isn't threatened by the fact that his largest customers — from Amazon Web Services and Microsoft to OpenAI — are trying to build their own chips.
Huang said Nvidia is improving at such a fast rate that he believes most of those companies will cancel their chip projects. He reiterated that building chips for AI is quite difficult. "If it was that easy, gosh, I don't know why I am working so hard," he said.
At the same time, Huang revealed a "super clever strategy." Some firms have approached Nvidia about using its networking technology to connect Nvidia GPUs with their own chips. These firms told Nvidia that if they can use its networking with their own chips, "we will buy everything else from you," meaning they would standardize their data centers with Nvidia equipment.
Companies are expected to spend $300 billion this year alone on AI infrastructure. The total AI chip market is projected to exceed $500 billion by 2028.
Nvidia is playing a masterful long-term game. Even as AMD mounts its most credible hardware challenge in years and customers try building their own chips, Nvidia is building platform control over how AI infrastructure is accessed, managed, and scaled.
This is the classic tech platform playbook. AWS didn't just win cloud computing by building better servers — it won by making cloud services essential. Nvidia is positioning itself as the indispensable layer between AI developers and the compute they need.
We're watching three different competitive strategies collide: AMD fighting the hardware war, customers trying to build their own chips, and Nvidia trying to win the platform war. The winner may determine who shapes artificial intelligence over the next decade.
Which image is real? |
🤔 Your thought process:
Selected Image 1 (Left):
“The randomness of real life vs the uniformity of AI generation.”
“Apparently AI assumed that cafes and restaurants should be crowded, and this premise is realized superficially. ”
Selected Image 2 (Right):
“Nuts. Looked like the boats were on different levels on the same body of water in the real one.”
“Very difficult to tell...thought the short view would seem easier to the AI engine to create.....”
💭 A poll before you go
Who wins the AI infrastructure battle? |
Thanks for reading today’s edition of The Deep View!
We’ll see you in the next one.
P.S. Enjoyed reading? Take The Deep View with you on the go! We’ve got exclusive, in-depth interviews for you on The Deep View: Conversations podcast every Tuesday morning. Subscribe here!
If you want to get in front of an audience of 450,000+ developers, business leaders and tech enthusiasts, get in touch with us here.