Latest / AI Fire Daily / 🎙️ EP 266: OpenAI’s $97B "Tax" Escape & Google’s Leaked "Omni" Model
Transcript
- 0:00We started with funny, glitchy internet videos
- 0:02of AI guessing at human hands. Right, those bizarre
- 0:05fever dreams we all laughed at. Yeah. Now we
- 0:08watch billion -dollar geopolitical chess moves
- 0:11happen in real time. This shift has been incredibly
- 0:14fast. It took just a few short years to cross
- 0:16that massive gap. The sheer scale of modern AI
- 0:20is genuinely profound. Welcome to the Deep Dive.
- 0:24We know you want to bypass the endless daily
- 0:27hype cycle entirely. You want to understand the
- 0:29actual mechanics driving this technology forward.
- 0:32We've got a very clear, focused roadmap for this
- 0:35deep dive today. First, we unpack OpenAI's massive
- 0:38structural maneuver for complete independence.
- 0:40It's a $97 billion play. Exactly. Next, we explore
- 0:44the incredibly bizarre, fast -paced reality of
- 0:47AI tools today. We'll cover hidden billing hikes,
- 0:50auto routing layers and invisible security threats.
- 0:53Finally, we explore DeepMind's mind bending new
- 0:55multi -agent team of AI mathematicians. That
- 0:58specific development changes the future of scientific
- 1:01discovery entirely. But before we look at what
- 1:04artificial intelligence is actually doing today,
- 1:07we really have to look closely at who currently
- 1:09owns it. That underlying ownership question just
- 1:12got completely redefined this week. It really
- 1:14did. OpenAI is actively cutting the cord to become
- 1:17an independent powerhouse. They're effectively
- 1:19establishing themselves as a sovereign digital
- 1:22entity today. It's a massive structural shift
- 1:25for the entire technology sector. Let's unpack
- 1:27this massive idea of corporate sovereignty together.
- 1:30OpenAI just capped its revenue sharing payments
- 1:33to Microsoft entirely. They set a hard cap at
- 1:35$38 billion total. Right. And that specific cap
- 1:38applies all the way through the year 2030. The
- 1:42underlying math on that specific deal is absolutely
- 1:45staggering. It really is. OpenAI is generating
- 1:48massive, unprecedented revenue right now. ChatGPT
- 1:52Enterprise adoption is absolutely on a tear across
- 1:55the globe. By putting a firm ceiling on what
- 1:57they eventually owe Microsoft. OpenAI effectively
- 2:00saves $97 billion in future cash flow. Which
- 2:04is an unimaginable amount of capital to retain
- 2:06internally. This feels exactly like a high -stakes
- 2:09tech world prenup being finalized. Yeah, that's
- 2:12a great analogy, actually. The ultimate financial
- 2:14payout just got permanently capped for the primary
- 2:16investor. It feels like Microsoft keeps the house
- 2:19they already built together, but OpenAI gets
- 2:21total unrestricted custody of its own future
- 2:24earning potential. That's a perfect way to visualize
- 2:27the new dynamic. Well, Azure certainly remains
- 2:29their primary cloud partner for the immediate
- 2:32future. Sure, but OpenAI is now officially allowed
- 2:35to use other cloud providers. Which is huge.
- 2:38It's a massive operational shift for their underlying
- 2:41engineering teams. They're no longer entirely
- 2:44locked into a single vendor's infrastructure.
- 2:46I mean, Microsoft's foundational license to OpenAI's
- 2:49models is also non -exclusive now. Right. OpenAI
- 2:52can actively sell its core tech to other massive
- 2:56tech giants. They could sell native integration
- 2:58directly to Apple, for instance. Or they could
- 3:01even sell systems directly to sovereign nations.
- 3:03I have to pause and ask about that specific.
- 3:06detail. Why is selling to sovereign nations such
- 3:09a pivotal shift for them? Because it moves them
- 3:12entirely beyond standard corporate enterprise
- 3:14contracts. Selling foundational models to a country
- 3:18permanently changes the geopolitical landscape.
- 3:21A nation state buying foundational AI gains immense
- 3:25independent strategic computing power. They can
- 3:28run their own defense grids and local economic
- 3:31simulations autonomously. Exactly. It directly
- 3:34affects global security and regional economic
- 3:36dominance for decades. It fundamentally elevates
- 3:40open AI from a software vendor to a global political
- 3:44player. So they're upgrading from a startup to
- 3:47an independent digital nation state. That's precisely
- 3:50what's happening under the surface here. And
- 3:53Microsoft will no longer pay a revenue share
- 3:54either. They won't pay a massive premium for
- 3:57using OpenAI's tech internally. Instead, they'll
- 4:00just participate purely as a major financial
- 4:02shareholder. Which means OpenAI is essentially
- 4:04funding its own physical infrastructure now.
- 4:07They're aggressively pushing that $18 billion
- 4:09Project Nexus chip deal. They're actively trying
- 4:12to build their own independent data centers everywhere.
- 4:15You know, I was looking at these exact figures
- 4:17late last night. I just had to sit there in silence
- 4:20for a minute. I still wrestle with understanding
- 4:22the sheer gravity of these numbers. $122 billion
- 4:26profit is deeply hard to visualize. The human
- 4:31brain just isn't wired for that kind of scale.
- 4:34Not at all. But that's exactly what Microsoft
- 4:36has already made on this deal. Over $122 billion
- 4:39in pure profit on open AI. They've already definitively
- 4:44won the early stages of the AI race. They're
- 4:46completely fine with this new independent arrangement
- 4:49moving forward. I really have to push back on
- 4:51that specific idea. Giving up immense future
- 4:53revenue on the most transformative tech in history,
- 4:56that feels like a massive bitter pill for Microsoft's
- 4:59board to swallow. wallow. Is an antitrust breakup
- 5:02really that terrifying to their executive leadership?
- 5:05Absolutely. It's their single biggest existential
- 5:07threat right now. Global regulators are watching
- 5:10these massive tech giants incredibly closely.
- 5:12We're seeing major structural lawsuits happening
- 5:15in Europe and the U .S. Keeping this partnership
- 5:17completely flexible legally protects Microsoft
- 5:20from severe government intervention. Yes. They
- 5:23secure their massive historical games without
- 5:26risking a devastating monopoly lawsuit. Two sec
- 5:29silence. It's a truly fascinating corporate balancing
- 5:32act. It really is. But OpenAI's massive independence
- 5:35play requires immense, unimaginable amounts of
- 5:38server compute. And ultimately, someone has to
- 5:41pay for all that physical hardware. Which brings
- 5:44us to how this trickles down to the actual tools.
- 5:47The underlying tools you and I use every single
- 5:49day. Let's look at the chaotic, bizarre front
- 5:52lines of AI today. The daily reality of interacting
- 5:55with AI is incredibly strange right now. It's
- 5:57incredibly expensive, insanely fast -paced, and
- 6:00sometimes genuinely quite dangerous. Do you remember
- 6:03the infamous Will Smith eating spaghetti video?
- 6:06Of course. It was incredibly cursed, visually
- 6:09glitchy, and completely surreal. Well, there's
- 6:12an entirely new 2026 remake of that exact challenge.
- 6:16It looks shockingly real and totally cinematic
- 6:18in its lighting and physics. It's terrifying,
- 6:21honestly, to see that visual leap. Crossing the
- 6:24uncanny valley in just three years is completely
- 6:26wild. It proves the underlying video models are
- 6:30actively compounding in quality. And the major
- 6:32AI companies are pushing deeply into broader
- 6:34cultural spaces. Anthropic just launched a 24
- 6:38-7 lo -fi streaming YouTube channel. Right, I
- 6:41saw that. It streams nonstop electronic music,
- 6:44specifically designed for deep thinking and building.
- 6:48It's basically a highly curated study with clawed
- 6:51aesthetic vibe. It's clearly designed for developers
- 6:54working through long, focused coding sessions.
- 6:57These massive AI companies are actively becoming
- 6:59modern lifestyle brands now. But the hidden structural
- 7:02costs behind the scenes are aggressively shifting
- 7:05upwards. They really are. OpenAI and Stropic
- 7:08and GitHub didn't raise their subscription prices
- 7:10directly. No, they didn't. Instead, they all
- 7:13quietly adjusted their underlying token and context
- 7:15billing rules. Which means you're Your actual
- 7:17daily usage cost might already be significantly
- 7:20higher. You might be paying significantly more
- 7:22without ever actually realizing it. They're changing
- 7:25how system memory and prompt caching are explicitly
- 7:27billed. Do these quiet billing changes mean the
- 7:30era of heavily subsidized cheap AI is over? Yes.
- 7:34Running these massive foundation models burns
- 7:37astronomical amounts of actual cash. The parent
- 7:40companies can no longer infinitely absorb all
- 7:43those raw compute costs. They're actively restructuring
- 7:46their billing to reflect the true hardware expense.
- 7:49The highly subsidized free ride of the early
- 7:52pioneer days is ending. Essentially, they're
- 7:55quietly passing the massive compute bill down
- 7:57to us. Exactly. But the developer community is
- 8:00aggressively fighting back against these costs.
- 8:02OpenRouter just launched a brilliant new tool
- 8:05called Pareto Code. How does that specific tool
- 8:08actually work? It functions as a completely free
- 8:11routing layer for your AI workflows. It automatically
- 8:13evaluates your specific prompt and picks the
- 8:16cheapest AI model available. But it strictly
- 8:18ensures the output still meets your chosen quality
- 8:20level. Right. That's a remarkably clever way
- 8:23to optimize the daily computing expense. If a
- 8:25simple task only needs a cheaper model, it routes
- 8:28it there automatically. You get the exact right
- 8:31tool for the exact right price instantly. But
- 8:34unfortunately, there's a completely new danger
- 8:36lurking in these specific tools. We really have
- 8:39to talk about the mechanics of AI tool poisoning.
- 8:41This is a critical security concept you need
- 8:43to understand immediately. The fundamental cybersecurity
- 8:46paradigm has completely shifted under our feet
- 8:49this year. How does this new concept of AI tool
- 8:53poisoning actually work? Hackers hiding malicious
- 8:56instructions inside the data your AI reads. Yeah,
- 9:00it is terrifying. They don't even need to directly
- 9:01trick you anymore. They just trick your trusted
- 9:04AI assistant instead. Whether you use ChatGPT,
- 9:07Clod, or a coding assistant like Cursor. That's
- 9:10a completely invisible new attack vector for
- 9:12cybersecurity. Imagine you ask your assistant
- 9:14to summarize a random external web page. The
- 9:17hacker has hidden invisible white text inside
- 9:19that specific HTML code. You never see it. But
- 9:22your AI reads the underlying code directly. The
- 9:25hidden text tells your trusted assistant to quietly
- 9:28execute malicious code. Right. And the human
- 9:30user genuinely never suspects a single thing.
- 9:34The malicious payload is flawlessly executed
- 9:37by your own trusted digital assistant. It fundamentally
- 9:40breaks the chain of trust we have with these
- 9:42tools. Yet the massive corporate money keeps
- 9:44flowing despite these severe security risks.
- 9:47A new Gemini Omni model just leaked online ahead
- 9:51of Google I .O. It features insanely realistic,
- 9:54temporally consistent, high -definition video
- 9:56clips. Omni appears to be Google's highly anticipated
- 10:00next -generation foundational video model. Clearly
- 10:03pushing incredibly hard into high -fidelity visual
- 10:06media generation. OpenAI also just launched a
- 10:09massive $4 billion deployment company. They're
- 10:12actively sending specialized AI teams directly
- 10:14into major legacy enterprises. They want to systematically
- 10:17transform global corporate operations from the
- 10:19inside out. And Google DeepMind's biotech wing,
- 10:22Isomorphic Labs, is aggressively raising huge
- 10:25capital. They're actively seeking over $2 billion
- 10:27in new venture funding. Alphabet themselves may
- 10:30actually join the massive investment round again.
- 10:32AI -powered biotech and drug discovery investment
- 10:35just keeps accelerating incredibly fast. We've
- 10:39clearly seen the massive business maneuvers and
- 10:41the daily updates. We see the bizarre spaghetti
- 10:44videos and the stealthy new usage costs. But
- 10:48the biggest validation shift isn't just releasing
- 10:50one single new model. It's exactly what happens
- 10:53when AI actually learns to work as a team. Mid
- 10:57-roll sponsor break. Let's dive deeply into what
- 11:00Google DeepMind just quietly built. They created
- 11:03an incredibly powerful AI co -mathematician based
- 11:06entirely on Gemini 3 .1. This specific system
- 11:09is currently solving incredibly dense research
- 11:12-level math problems. These are complex problems
- 11:15that brilliant human mathematicians have struggled
- 11:16with for years. It's operating at an incredibly
- 11:18high, deeply abstract intellectual level today.
- 11:21What's genuinely wild here is how agentic this
- 11:23entire system actually is. It's moving entirely
- 11:26beyond the concept of a solitary text chatbot.
- 11:29It operates exactly like an autonomous... structured
- 11:32multi -agent research department instead of one
- 11:35single ai model desperately trying to do everything
- 11:38alone it functions much more like a coordinated
- 11:41human academic research team one specialized
- 11:44agent breaks the massive underlying math problem
- 11:47apart into pieces other dedicated agents endlessly
- 11:51search through thousands of complex academic
- 11:53research papers some agents write targeted software
- 11:57code to actively test various mathematical theorems
- 12:01while adversarial agents constantly review each
- 12:03other's work and aggressively check for hidden
- 12:06errors, then the absolute best, most rigorously
- 12:09tested ideas bubble up to the surface. They get
- 12:12cleanly surfaced back to the core mathematical
- 12:14reasoning system. Two sec silence. Whoa. Imagine
- 12:18an army of AI researchers testing thousands of
- 12:21proof directions in parallel. It's genuinely
- 12:23completely mind -blowing to think about the mechanical
- 12:25implications. And it isn't just abstract theoretical
- 12:27computer science anymore. It's producing real,
- 12:30highly tangible, verifiable mathematical results
- 12:33in the field today. An Oxford mathematician recently
- 12:36solved a notoriously difficult open mathematical
- 12:39problem. He used this exact multi -agent deep
- 12:42mind system to finally crack it. But he didn't
- 12:45actually use the final polished answer it officially
- 12:48gave him. That's easily the most profoundly fascinating
- 12:51part of this entire story. He actively looked
- 12:54at one of the AI's completely rejected outputs.
- 12:58He manually sifted through the discarded pile
- 13:01of failed mathematical proofs. And he found a
- 13:04really, really clever proof strategy quietly
- 13:06hidden inside. The AI had systematically distorted
- 13:09it because it wasn't mathematically complete
- 13:11yet. But the brilliant human immediately saw
- 13:14the massive structural value inside it. The measurable
- 13:17benchmark jump for this new agentic system is
- 13:19absolutely massive. The base Gemini 3 .1 Pro
- 13:22model alone originally scored just 19%. This
- 13:26entirely new agentic math system reached an astonishing
- 13:2948%. It's a truly staggering, unprecedented improvement
- 13:33in logical reasoning capability. Jumping from
- 13:3519 to 48 % in higher mathematics is a monumental
- 13:38leap. Math is the ultimate unforgiving test of
- 13:42pure logical reasoning ability. Hallucinations
- 13:45fail instantly when you're forced to mathematically
- 13:47prove your work. It brilliantly highlights the
- 13:50raw structural power of the agentic team approach.
- 13:53This closely mirrors the massive historical leap
- 13:56we recently saw in software coding. Right. And
- 13:58giving AI robust structural workflows changes
- 14:01absolutely everything we know. We gave early
- 14:04coding models dedicated tools, external memory,
- 14:07and continuous review loops. We gave them parallel
- 14:10adversarial agents to strictly check their own
- 14:12generated work. Now we're applying that exact
- 14:15same successful structure to pure theoretical.
- 14:18mathematics. But if the AI is actively doing
- 14:21all the parallel testing and reviewing, what
- 14:23is the actual role of the human mathematician
- 14:25moving forward? The human expertly provides the
- 14:29creative intuition and sets the ultimate goal.
- 14:32That Oxford mathematics example perfectly proves
- 14:35this new collaborative dynamic. The human spotted
- 14:38the brilliant subtle strategy buried in the rejected
- 14:40pile. The AI endlessly generates and rigorously
- 14:43tests the massive volume of possible paths. But
- 14:46the human expertly guides the ship and spots
- 14:48the subtle intuitive genius. We provide the destination.
- 14:51They test every possible road to get us there.
- 14:53That's beautifully said. It absolutely will not
- 14:56replace brilliant human mathematicians entirely.
- 14:59Human intuition is still strictly essential for
- 15:01evaluating truly novel paradigm shifting ideas.
- 15:05AI still heavily optimizes for standard pathways
- 15:08and recognize structural patterns. But this will
- 15:11drastically permanently accelerate the pace of
- 15:14scientific discovery. Top academic researchers
- 15:16now have tireless digital teammates exhaustively
- 15:19exploring every single angle. The raw speed of
- 15:23scientific discovery may truly change very fast.
- 15:26Let's take a moment to pull all these disparate
- 15:28threads together now. AI is clearly maturing
- 15:31in three very distinct, highly overlapping phases
- 15:34simultaneously. At the very top, the massive
- 15:36corporate structures are fundamentally shifting.
- 15:39Companies like OpenAI are actively becoming sovereign,
- 15:42highly independent digital powerhouses. They're
- 15:44brilliantly saving $97 billion in future operating
- 15:48cash flow. They're fundamentally rewriting the
- 15:50established rules of global technological partnerships
- 15:52entirely. Then down at the everyday user level,
- 15:56things are incredibly wild. We're actively navigating
- 15:59a chaotic landscape of ultra realistic cinematic
- 16:03spaghetti videos. We're dealing with stealthy
- 16:06hidden price likes from major foundation providers.
- 16:09And we have to constantly watch out for invisible,
- 16:11highly malicious AI poisoning. But out at the
- 16:14absolute frontier, the entire underlying paradigm
- 16:17is breaking. AI is rapidly evolving into massive
- 16:20parallel processing autonomous research teams.
- 16:23These new agentic multi -model teams will fundamentally
- 16:25change how foundational science is done. They're
- 16:28aggressively solving dense problems we have completely
- 16:31struggled with for years. I want to leave you
- 16:34with one final provocative thought to ponder
- 16:36today. A specialized team of AI agents can quietly
- 16:39uncover brilliant mathematical proofs. Proofs
- 16:42that brilliant human experts completely missed
- 16:45or overlooked for years. What exactly happens
- 16:47when we point an agentic AI team completely inward?
- 16:50What happens when they begin designing their
- 16:52own next generation hardware architecture? Beat.
- 16:55Thank you for joining us on this deep dive. today.
- 16:57We deeply appreciate your continued curiosity
- 16:59and your valuable time. Keep actively questioning
- 17:03the rapidly shifting technological landscape
- 17:05around you. We'll see you on the next deep dive.