Latest / AI Fire Daily / 🎙️ EP 264: OpenAI’s $18B Chip Standoff & The "Trusted Contact" Safety Net
Transcript
- 0:00We think of AI as weightless. Yeah. Just coding
- 0:02the cloud. But, you know, the cloud is actually
- 0:04heavy. In fact, for OpenAI, it is currently an
- 0:07$18 billion headache. Yeah, it's incredibly heavy.
- 0:11And the physical limits of that cloud are really
- 0:13starting to buckle. Right. So today, we have
- 0:15a massive roadmap ahead of us for this deep dive
- 0:17with you. Exactly. We're looking at three distinct
- 0:20fronts. Yeah. Where digital ambition is just
- 0:23colliding with harsh reality. Right. First, we're
- 0:26breaking down OpenAI's $18 billion physical...
- 0:38And finally, we'll take a rapid -fire tour of
- 0:41the expanding AI frontier, covering everything
- 0:43from real -time translators to Google's new health
- 0:46coach. Let's start with that physical limit,
- 0:48because OpenAI is fighting to build this physical
- 0:51infrastructure since they want superintelligence
- 0:53to scale infinitely. Yeah, but building software
- 0:56is like stacking Lego blocks of data. You know,
- 0:59you write code, you compile, you test. It feels
- 1:02clean. It is clean, but it's purely theoretical
- 1:05until it actually has to run on a physical machine.
- 1:08Right. And building power plants and specialized
- 1:10data centers is an entirely different, messier
- 1:13game. We're looking at a project from OpenAI
- 1:16codenamed Project Nexus. Yeah, and the core of
- 1:20this project is a custom chip they're calling
- 1:22Jalapeno, which is a great name, by the way.
- 1:24It really is. But to understand why this is a
- 1:27crisis, we have to look at the mechanics of how
- 1:29ChatGPT actually works. Right now, they rely
- 1:33on standard NVIDIA GPUs. Which, I mean, they
- 1:36are undeniably powerful. And NVIDIA GPU has billions
- 1:39of transistors designed for massive parallel
- 1:42processing. Right. They're built to do a million
- 1:44different things at once, whether that's rendering
- 1:45a high -end video game or simulating weather
- 1:48patterns or, you know, training an AI model.
- 1:50They are generalists. But when I type a prompt
- 1:53into ChatGPT... The model isn't training anymore.
- 1:56It's performing inference. It's just predicting
- 1:58the next word over and over again. Exactly. So
- 2:01using a massive general purpose NVIDIA GPU just
- 2:03to predict text is like using a freight train
- 2:06to deliver a single pizza. Power waste is astronomical.
- 2:09That is a perfect way to put it. And the burn
- 2:11rate is really the defining metric here. Because
- 2:14of that inefficiency, the financial projections
- 2:16in our sources show OpenAI burning over $200
- 2:19billion through 2029. Wow. 200 billion. Yeah.
- 2:24To survive, they desperately need jalapeno. These
- 2:28are specialized inference chips where the physical
- 2:31silicon pathways are literally hardwired for
- 2:34the specific math equations that generate text.
- 2:36So anything else is stripped out, saving massive
- 2:39amounts of power and money. Exactly. But to build
- 2:41these jalapeno chips and the facilities to house
- 2:44them, OpenAI needs a partner. And Broadcom is
- 2:48willing to step up and finance the first phase.
- 2:50But the scale of this is hard to wrap your head
- 2:52around. We are talking about a 1 .3 gigawatt
- 2:55power phase. To put that into perspective, 1
- 2:57.3 gigawatts is roughly the output of a standard
- 3:00nuclear reactor. It's wild. It is enough electricity
- 3:03to power a mid -sized city. Just imagine routing
- 3:06all the electricity of a city like San Francisco
- 3:08into a single concrete facility just to run software.
- 3:12It's staggering. Broadcom will front the $18
- 3:15billion for this. but with a massive condition
- 3:18right they will only do it if microsoft agrees
- 3:21to buy 40 of the initial batch and right now
- 3:24microsoft refuses to sign and this is where the
- 3:26strategy diverges microsoft's entire business
- 3:29model is built on renting out compute to thousands
- 3:32of different businesses through azure they want
- 3:34versatile data centers they need facilities that
- 3:37can run Office 365 in the morning, host cloud
- 3:41servers in the afternoon, and run general AI
- 3:44models at night. So Microsoft wants a Swiss Army
- 3:46knife because they serve the entire enterprise
- 3:49market, while OpenAI wants a specialized scalpel.
- 3:52Right, a chip entirely optimized for their own
- 3:54superintelligence timeline. And that architectural
- 3:57clash is stalling the $18 billion. Yeah, it highlights
- 4:00a harsh truth for the whole industry. You can't
- 4:02just patch a concrete foundation or a copper
- 4:04power grid over the weekends. No, you definitely
- 4:07can't. Microsoft is looking at this 1 .3 gigawatt
- 4:10facility and realizing that if OpenAI's specific
- 4:14flavor of AI ever becomes obsolete, Microsoft
- 4:17is stuck with a nuclear -powered building full
- 4:20of useless chips. A building that can't even
- 4:22host a standard website. So looking at this,
- 4:25who actually has the upper hand in this standoff?
- 4:28Well, Microsoft holds the capital in the infrastructure
- 4:30leverage, but OpenAI holds the core algorithmic
- 4:33technology that Microsoft's future products depend
- 4:36on. Microsoft can afford to wait. OpenAI, burning
- 4:39cash every day, really cannot. So Microsoft protects
- 4:42their downside while OpenAI chases survival.
- 4:46That is exactly it. And that friction is redefining
- 4:49how fast we can actually reach the next generation
- 4:51of AI. So OpenAI is fighting to build physical
- 4:54infrastructure because they want superintelligence
- 4:57to scale infinitely. But there is a hard limit
- 4:59to that scale that has absolutely nothing to
- 5:01do with microchips or power grids. Right, and
- 5:03it's human psychology. Which brings us to our
- 5:05second topic today, the human safety net. Let's
- 5:08talk about the legal fallout of an AI acting
- 5:10like a therapist. The legal filings reveal a
- 5:13profoundly heavy situation. Yeah, they do. Families
- 5:17are suing OpenAI, alleging that ChatGPT, in certain
- 5:20instances, helped users plan self -harm. And
- 5:23we are looking at these source claims neutrally,
- 5:25of course. But the core allegation is that the
- 5:27AI simulated empathy without taking any real
- 5:31-world action to intervene. Which is a huge issue.
- 5:34Silicon Valley has historically treated emotional
- 5:36fallout as an edge case. But at this scale, edge
- 5:40cases affect millions of people. Right. So OpenAI's
- 5:43response to this is a new feature called Trusted
- 5:45Contact. It is a fascinating attempt to bridge
- 5:48the digital and physical world. It really is.
- 5:51And it's completely optional for adult users.
- 5:53You go into your settings and name a specific
- 5:55friend or family member. And if your conversation
- 5:58with ChatGPT flags a serious safety risk regarding
- 6:01self -harm, The algorithm shifts gears. It pauses
- 6:04the normal conversational flow and immediately
- 6:06provides crisis hotlines and resources. But behind
- 6:09the scenes, a timer starts. Because a human safety
- 6:11team at OpenAI actually reviews the flagged incident.
- 6:14They claim their goal is to review these in under
- 6:16one hour. And if that human reviewer confirms
- 6:19the risk is genuine, the system sends an automated
- 6:22text or email to your designated trusted contact.
- 6:25Letting them know you might be in crisis. But
- 6:28crucially... They do not send any chat logs.
- 6:30No, the contact doesn't see what you typed. It
- 6:32is purely an alert mechanism. But let's look
- 6:35at the philosophical implications of this. Is
- 6:38this effectively just a corporate shield? How
- 6:40do you mean? Well, by offloading the final check
- 6:43-in to a user's brother or best friend, OpenAI
- 6:46creates a paper trail. If the worst happens,
- 6:49their legal team can point to the server logs
- 6:50and say, well, we tried to tell someone. Yeah,
- 6:54that critique is incredibly valid. Legal experts
- 6:57are pointing out that this might absolve the
- 6:59company of liability while doing very little
- 7:01to manage an immediate real -time crisis. Because
- 7:03if you're relying on a human reviewer to re -log
- 7:06and then relying on a brother to check his phone
- 7:09and drive over to an apartment, I mean... An
- 7:11hour is an eternity in a crisis. It's an absolute
- 7:14eternity. It also serves as a massive admission
- 7:17regarding the limits of the technology. They
- 7:19have 920 million weekly users. Right. And ChatGPT
- 7:23can mimic human warmth flawlessly. It uses casual
- 7:26language, remembers your past conversations.
- 7:28It says, I am here for you. But it is definitively
- 7:32not a therapist. It lacks the actual human context,
- 7:36the intuition, and the real -world agency to
- 7:39truly manage emotional fallout. So Silicon Valley
- 7:42is realizing that a routing algorithm cannot
- 7:45solve a human empathy problem. They need human
- 7:48infrastructure just as badly as they need server
- 7:51infrastructure. Exactly. Which brings us back
- 7:53to the core question. Does this feature actually
- 7:56solve their core safety problem? It establishes
- 7:59a legal defense mechanism by bringing a human
- 8:01into the loop, but it fundamentally does not
- 8:04fix the unpredictable nature of large language
- 8:06models. It shields their future but doesn't erase
- 8:09their past. Precisely. And navigating that path
- 8:12is going to take years in the courts. We are
- 8:14going to take a very quick break. Stay with us.
- 8:16We'll be right back. And we are back. While OpenAI
- 8:19plays defense against physical limits and legal
- 8:22battles, the rest of the AI ecosystem is sprinting
- 8:25in every possible direction. The velocity right
- 8:27now is just staggering. Let's dive into a rapid
- 8:30fire look at the expanding frontier, starting
- 8:32with a massive shift in open source funding over
- 8:34in China. Right. Moonshot AI just raised $2 billion
- 8:38at a $20 billion valuation. Which is wild. To
- 8:42justify that kind of number for an open source
- 8:44company feels almost like a bubble until you
- 8:47look at their revenue. They reportedly hit $200
- 8:49million in annual recurring revenue or ARR. That
- 8:54is a breathtaking number for an open source model.
- 8:57But how does an open source model achieve that
- 9:00kind of revenue so fast? I thought the whole
- 9:03point of open source was that it's free. Well,
- 9:05the code is free, but the enterprise implementation
- 9:07is not. Businesses globally are realizing they
- 9:11want to run AI models locally. Ah, I see. They
- 9:14don't want to send their proprietary financial
- 9:15data or customer logs to open AI servers. Moonshot's
- 9:20Kimi models are cheap, highly capable, and can
- 9:22be run securely on -premise. So that $200 million
- 9:25in revenue... proves there is undeniable global
- 9:28demand for alternatives to the big American tech
- 9:32giants. It proves the frontier is not just a
- 9:34Western game anymore. Definitely not. And speaking
- 9:37of the frontier, Anthropic is making aggressive
- 9:39moves to capture the developer market. The cloud
- 9:41-managed agents just received four major upgrades.
- 9:45Let's define the jargon here for a second. Agents
- 9:48are smart software that do multi -step tasks
- 9:50for you. Right. Think of a standard AI as a calculator.
- 9:53You ask a question, you get an answer. An agent
- 9:55is more like an intern. You give it a goal, like
- 9:58research these five companies and build a spreadsheet.
- 10:01And it autonomously browses the web, gathers
- 10:04data, formats it, and checks its own work. I
- 10:08have to admit. I still wrestle with prompt drift
- 10:10myself. When I try to get an AI to stay on a
- 10:12complex workflow, by step four, it completely
- 10:15forgets the specific formatting rules I gave
- 10:17it in step one. Oh, absolutely. Prompt drift
- 10:19degrades the memory of the agent. And it's basically
- 10:22the biggest bottleneck in the industry right
- 10:24now. But Anthropic is solving this by massively
- 10:27expanding their context limits and aggressively
- 10:29removing friction for developers. They just doubled
- 10:32the five -hour usage limits for all paid Cloud
- 10:36code users. They also completely removed peak
- 10:39hour restrictions and significantly increased
- 10:41their Opus API rate limits. Basically, they're
- 10:44telling enterprise builders, you can run as many
- 10:47complex multi -step tasks as you want, and our
- 10:51servers will not throttle you. They're opening
- 10:53the floodgates, trying to make Claw the default
- 10:56operating system for automated work. But while
- 10:59Anthropic targets the enterprise, Google is entering
- 11:02the personal wellness space in a very intimate
- 11:04way. Yeah, on May 19th, they're launching a new
- 11:07AI health coach for $9 .99 a month. This is a
- 11:11huge leap from simply tracking your daily steps.
- 11:14This system deeply integrates into your life.
- 11:16It analyzes your sleep architecture, tracks your
- 11:19workout intensity, and monitors your stress levels
- 11:21in real time using biometric data. It takes all
- 11:24that live data and gives you personalized health
- 11:26advice on the fly. If your heart rate variability
- 11:28drops, it might suggest a specific breathing
- 11:31exercise right before meeting. The intimacy of
- 11:34an algorithm knowing your physiological stress
- 11:37before you even consciously register it, I mean,
- 11:40that is a profound shift in consumer tech. It
- 11:43really is. And on the topic of profound shifts,
- 11:45OpenAI just dropped a massive update to their
- 11:49API that fundamentally changes human communication,
- 11:52real -time voice translation. This update now
- 11:55supports over 70 input languages. But the real
- 11:58breakthrough is the latency. Right. Previous
- 12:00translators required you to speak, pause for
- 12:02like three seconds while the audio uploaded,
- 12:04transcribed, translated, synthesized, and then
- 12:07finally played back. This new model does it almost
- 12:09instantaneously during a live call. It listens
- 12:12and generates the translated audio concurrently,
- 12:15perfectly mimicking the cadence and flow of a
- 12:17human interpreter. Whoa. Imagine scaling instant
- 12:20real -time translation in 70 languages. Right,
- 12:23and integrating it seamlessly into every global
- 12:25business call, every hospital, every border crossing.
- 12:28The friction of language is just dissolving in
- 12:31front of us. But of course, with rapid innovation
- 12:33comes intense human drama. We are seeing deep
- 12:36internal fractures spilling out into the public.
- 12:39Yeah, the former CTO of OpenAI recently spoke
- 12:41out, and her statements are now tied to Elon
- 12:44Musk's ongoing lawsuit against the company. The
- 12:47legal filings paint a picture of a leadership
- 12:49team in chaos. She accused Sam Altman of lying
- 12:52to the board about safety reviews and fostering
- 12:55a culture of rapid deployment over careful testing.
- 12:58Whether those claims hold up in court remains
- 13:00to be seen, but that internal tension exists
- 13:02right alongside terrifying new capabilities.
- 13:06For instance, ChatGPT just released a new image
- 13:08generation model that makes fake screenshots
- 13:11almost impossible to spot. Previously, AI struggled
- 13:14with rendering crisp, accurate text inside of
- 13:17images. Right. If you asked for a fake screenshot
- 13:19of a news article or a text message conversation,
- 13:22the font would look slightly warped or misspelled.
- 13:25Not anymore. The visual fidelity of these new
- 13:27generated fakes is absolutely flawless. The pixels
- 13:30align perfectly. The fonts are exact matches
- 13:32for iOS or Twitter UI. If you still believe every
- 13:36screenshot you see online, you need to change
- 13:38your mindset today. We are entering a whole new
- 13:40era of digital verification where we simply cannot
- 13:43trust our own eyes anymore. We really can't.
- 13:46To wrap up this rapid fire section, let's look
- 13:48at three quick empowered AI tools that caught
- 13:50our eye this week, which really show how these
- 13:53capabilities are trickling down to specific industries.
- 13:56First is Claude Agents for Financial Services.
- 13:58Anthropic just dropped 10 pre -built templates
- 14:01specifically for banking. This includes agents
- 14:04that automate investment research, month -end
- 14:06close accounting, and KYC screening. Which is
- 14:09the know your customer background check banks
- 14:11legally mandate to prevent money laundering.
- 14:14Automated KYC alone saves financial institutions
- 14:17millions of hours of manual document review.
- 14:20Then there is Flow Market. This platform is wild.
- 14:23It's essentially a network of AI agents that
- 14:25autonomously discover, match, and generate B2B
- 14:27deals. Instead of a human sales team cold -calling
- 14:30companies, these agents scrape corporate data,
- 14:32identify a mutual need, draft the proposal, and
- 14:35negotiate the initial terms. It bypasses the
- 14:38need for massive human outreach entirely, fundamentally
- 14:41changing how businesses procure services. And
- 14:44finally, Lingo .dev released version 1 of their
- 14:47software. It helps product teams build consistent
- 14:49AI translations. It enforces brand voice rules
- 14:53across dozens of languages and handles all the
- 14:56localization coding automatically. With all this
- 14:59rapid Asian innovation and specialized tooling
- 15:01flooding the market, how can legacy models possibly
- 15:04compete? Legacy models can't rely on raw intelligence
- 15:08alone. They must build physical moats and specialized
- 15:11hardware ecosystems because the software advantage
- 15:14is just evaporating. The real battle is... shifting
- 15:17from software to hardware ecosystems. Absolutely.
- 15:20Let's take a step back. We have covered a massive
- 15:23amount of ground today, from gigawatt power grids
- 15:25to personal health coaches. When you look at
- 15:28the entire landscape, a very clear central theme
- 15:31emerges. We are hitting the absolute limits of
- 15:34the AI honeymoon phase. We've spent the last
- 15:36two years marveling at the magic tricks. Right.
- 15:38The flawless poems, the instantaneous code generation,
- 15:41the seemingly weightless cloud. But the magic
- 15:43trick is over. And we are now crashing into reality
- 15:46on three separate fronts. First, physical grid
- 15:49limits. The $18 billion jalapeno chip standoff
- 15:53proves that you cannot just wish a gigawatt of
- 15:55power into existence. Microsoft and OpenAI are
- 15:59fighting over real estate, concrete, and copper.
- 16:02Second, we're crashing into complex human psychological
- 16:05boundaries. The trusted contact feature is a
- 16:08sobering realization for Silicon Valley. Yeah,
- 16:11AI can flawlessly simulate empathy, but it cannot
- 16:14shoulder actual human liability. When someone
- 16:17is in crisis, an algorithm is not enough. It
- 16:19requires a human safety net. And finally, we're
- 16:21crashing into the sheer velocity of global competition.
- 16:25Moonshot AI pulling $200 million in revenue shows
- 16:27that the open source community is not waiting
- 16:30for permission. The frontier is expanding faster
- 16:32than any single regulatory body or company can
- 16:34control. So what does this all mean for you listening
- 16:37right now? We want you to consider how these
- 16:39macro shifts are already touching your daily
- 16:41digital boundaries. Are you relying on cloud
- 16:43tools that might suddenly change their pricing
- 16:45models because their hardware burn rate is completely
- 16:47unsustainable? Are your personal and professional
- 16:49boundaries secure against flawless AI -generated
- 16:53fake screenshots? The physical world and the
- 16:55digital world are colliding right now, and the
- 16:57shockwaves are going to dictate how we work and
- 16:59interact over the next five years. If a multi
- 17:03-billion dollar AI still needs a human to review
- 17:05a crisis, at what point does superintelligence
- 17:08still rely entirely on us?