Latest / Deep Values: Value Investing Deep Dives / Pichai’s Google 2027 Enterprise AI Singularity Prediction
Transcript
- 0:00So every single week, the CEO of a $2 trillion company blocks off
- 0:05a full hour on his calendar, right?
- 0:07And you know, he doesn't spend it looking at quarterly earnings
- 0:11or reviewing marketing campaigns or even analyzing market share.
- 0:15You'd think you would, but no. No, exactly.
- 0:17Sundar Pichai literally sits down to manually micromanage TPU
- 0:22allocations across alphabets internal product teams.
- 0:25Which is just wild to think about.
- 0:27I mean, compute has effectively replaced capital has like, the
- 0:31most precious resource on the planet right now.
- 0:32Yeah, when the CEO of Google is, you know, down in the weeds,
- 0:37rationing silicon chips to individual engineering pods, it
- 0:40signals A fundamental shift in how the whole global economy is
- 0:43operating. Totally and welcome to the deep
- 0:46dive everyone. Today we are examining an
- 0:48incredibly revealing hour long interview from April 2026 on the
- 0:52Stripe Channel where Pichai sits down with John Collision and E
- 0:55LED Gill. Right.
- 0:56And it's important to note this is not your standard rehearsed
- 1:00PR circuit. No, not at all.
- 1:02It's a highly granular, really technical breakdown of his
- 1:05decade long tenure and we're going to use it to decode
- 1:09Alphabets staggering $175 billion to $185 billion capital
- 1:15expenditure for this year. I mean that number is just is
- 1:18hard to even wrap your head around.
- 1:19It really is. So we are going to look closely
- 1:22at the underlying mechanics driving that number.
- 1:24We'll examine the physical constraints currently choking
- 1:27the AI industry, the architectural philosophy behind
- 1:30their models, and you know why Pichai is officially pointing to
- 1:342027 as the timeline for an enterprise AI singularity.
- 1:38OK, let's untack this, because to understand where they're
- 1:41heading, we really have to look at the history of Pichai
- 1:43outlines here. Yeah, the context is everything.
- 1:46Right, because for the last few years, the dominant narrative
- 1:49has been, you know, Google invented the core technology,
- 1:52the transformer architecture, but somehow they just got beat
- 1:54to the punch like a massive luxury cruise ship that just
- 1:58can't turn as fast as a startup speed boat.
- 2:01Heard that cruise ship analogy so many times.
- 2:03Right, but I've always suspected that part of this narrative
- 2:06stems from a total misunderstanding of why the
- 2:09transformer was actually built in the first place.
- 2:11That is such a critical point. I mean, the Transformer paper
- 2:14wasn't just like a random theoretical academic exercise.
- 2:18Google engineered that architecture to solve a massive,
- 2:21highly specific infrastructure problem.
- 2:24They needed a way to power speech recognition and machine
- 2:27translation for an ecosystem of over 2 billion users, and they
- 2:31had to do it in real time 2. Billion users.
- 2:34Right, so the architecture was designed to parallelize
- 2:37processing at just an unprecedented scale.
- 2:40And it wasn't as if they just, you know, put it on a shelf and
- 2:42forgot about it. They actually integrated
- 2:45language models into search years ago with systems like Bert
- 2:49and MU. Yeah, people forget about those.
- 2:52They do, but from an outsider's perspective, those felt like
- 2:55back end optimizations, right? The public didn't get that
- 2:58visceral, conversational chatbot experience early on.
- 3:01No, they didn't. And you'd really think the
- 3:02pressure from shareholders to just, you know, show a shiny new
- 3:05product would have overridden their internal engineering
- 3:08caution. It's a super tempting business
- 3:10move for sure, but let's look at the mechanics of why they
- 3:13actually held the conversational models back.
- 3:16Long before the public saw any external chat bots, Google had
- 3:20Lambda running internally. Oh man, Lambda.
- 3:23I have to bring this up because the internal prototypes were so
- 3:26good that some engineers controversially thought it
- 3:31possessed actual self-awareness. Yes, the sentient AI drama.
- 3:35Like an engineer actually got fired over that whole debacle,
- 3:38so it was clearly highly capable.
- 3:40It was incredibly capable, but the issue is that early
- 3:42conversational models fundamentally lacked robust end
- 3:46to end reinforcement learning from human feedback or RLHF.
- 3:50Which, just to clarify for everyone, is much more than just
- 3:53a simple content filter you slap on top of a system, right?
- 3:56Precisely. It's not just a filter.
- 3:58ROHF is an entirely secondary training phase.
- 4:01You train a separate reward model based on human
- 4:04preferences, and that reward model mathematically penalizes
- 4:07the primary AI for generating, you know, toxic, unhelpful, or
- 4:12hallucinated text. Oh, I see.
- 4:13Yeah, it actually alters the underlying probability
- 4:16distribution of the models outputs.
- 4:18So without that reward model deeply integrated and rigorously
- 4:21tested, the baseline AI was highly volatile.
- 4:25And I guess when you are processing millions of queries a
- 4:28year, a volatile output isn't just like a minor bug, it's a
- 4:32catastrophic brand risk. Exactly.
- 4:34I mean, a startup can release a beta product, generate a few
- 4:38wildly inaccurate or offensive responses, apologize, and just.
- 4:41Reiterate right, move fast and break things.
- 4:43Right, but if Google search starts outputting fundamentally
- 4:46flawed or dangerous instructions on a global scale, it impacts
- 4:49the broader information ecosystem.
- 4:51Yeah, the safety floor for them is just massively higher.
- 4:54Exactly. They required a level of
- 4:55deterministic control that the underlying math simply didn't
- 4:58support yet. OK, so that standard for
- 5:01deterministic control is why they waited.
- 5:04But what's really fascinating in this interview is how Pachai
- 5:07describes their current competitive stance now that the
- 5:10products are actually out there. Yeah, it's a huge shift.
- 5:12The primary metric they are optimizing for now isn't just
- 5:15parameter count or getting bigger, it's speed.
- 5:19He seems entirely fixated on the millisecond war.
- 5:22He really does. Which, if I'm being perfectly
- 5:24honest, feels a bit misplaced from a user perspective.
- 5:27Like, does speed really matter that much for AI?
- 5:29Well, I mean, if an AI agent is writing a complex Python script
- 5:34for me or analyzing A dense marketing plan, I genuinely
- 5:37don't mind waiting an extra 2 or 5 seconds.
- 5:40It's saving me hours of manual labor.
- 5:42Yeah, that is the standard way to look at latency for sure.
- 5:45But Pichai argues that low latency isn't really about human
- 5:49patients. OK, then what is it about?
- 5:51It's an indicator of a fundamentally sound underlying
- 5:53system architecture. He talks at length about pushing
- 5:56what's called the Pareto frontier.
- 5:58Right, the Pareto frontier. That's the concept of finding
- 6:00the perfect balance or the mathematical limit of
- 6:02efficiency, right? Right.
- 6:04In machine learning, the Pareto frontier represents the absolute
- 6:07boundary where you cannot improve one variable without
- 6:11degrading another. So you are constantly balancing
- 6:14inference speed, model capability, and compute cost.
- 6:18If your system operates with incredibly low latency, it means
- 6:22your hardware and software are communicating with near perfect
- 6:25optimization. And they actually enforce this
- 6:28internally, don't they, through what they call a latency budget?
- 6:30Yes, which is fascinating. It operates almost like a like a
- 6:34corporate tax on engineering. That's a great way to put it.
- 6:36Yeah, like if a development team optimizes a piece of code and
- 6:40manages to save 3 milliseconds of delay in a process, they
- 6:43aren't permitted to use all three of those milliseconds to
- 6:46add a heavy new feature. Nope.
- 6:48They are strictly capped. Right, they're only allowed to
- 6:50use 1.5 milliseconds for the new feature.
- 6:54The remaining 1.5 milliseconds must be passed directly back to
- 6:57the user to make the fundamental platform feel faster.
- 7:00It's a brilliant structural forcing function for efficiency,
- 7:04and the result is that over the past five years, despite
- 7:08layering incredibly heavy AI neural networks in search,
- 7:12they've actually reduced overall latency by 30%.
- 7:15Wait, really? They lowered latency while
- 7:17adding. AI by 30%.
- 7:19That is wild, but I guess that underlying architectural speed
- 7:23isn't just a nice to have metric, it's literally the
- 7:26prerequisite for things like their Gemini flash models.
- 7:30Exactly. I mean with flash they've
- 7:31retained like 90% of the heavier pro models reasoning
- 7:35capabilities, but they are executing it with vastly
- 7:38superior speed and efficiency. And that's only possible through
- 7:41their full stack vertical integration.
- 7:44They're designing the TPU chips and the AI models in tandem, so
- 7:48the software is completely native to the silicon.
- 7:50Right. They're deeply synced, and that
- 7:51speed enables the next evolution Pichai discusses.
- 7:54Which brings us to a huge topic. This is where we move past the
- 7:57traditional search bar entirely, yes.
- 8:00The evolution of search. Yeah, for years now there's been
- 8:03this loud narrative claiming traditional search will be
- 8:05entirely replaced by AI chat bots within a decade.
- 8:08The death of the blue inks. Exactly.
- 8:10People think we're moving from a library card catalog straight to
- 8:13like a wizard who just hands you the answers.
- 8:16But is the search bar really dead?
- 8:19What Pichai frames it differently, he definitively
- 8:22says no, search won't be replaced.
- 8:25It's going to evolve into what he calls the agentic manager.
- 8:28An agentic manager. Right, it's not a 0 sum game.
- 8:31I mean, think about how YouTube continued to grow right
- 8:34alongside TikTok, or how Amazon thrived despite Google and
- 8:38Facebook. True, the market just expands.
- 8:40Exactly, if you look at the progression of search, it makes
- 8:43sense. We started with PC keywords,
- 8:46typing fragmented phrases into a box.
- 8:48Right, like weather. New York tomorrow.
- 8:50Exactly. Then we move to mobile voice
- 8:52search using actual natural language.
- 8:54The next phase isn't just retrieving information, it's
- 8:57executing asynchronous agentic workflows.
- 9:00OK, so think about how you currently plan a business trip.
- 9:03You are the orchestrator. You have one tab open for flight
- 9:06routing, another for hotel availability, maybe 1/3 for your
- 9:10corporate budget software, and a fourth for your calendar.
- 9:13It's a very manual process. Very.
- 9:15But Anagenic Manager doesn't just give you links to those
- 9:19platforms. It takes a single intent.
- 9:21Like, book me a compliant trip to London next week and it
- 9:24autonomously breaks it down to parallel subtasks.
- 9:27And it books everything and even adjusts the itinerary in the
- 9:31background if things change. And this this brings us back to
- 9:34why that strict latency budget is so critical.
- 9:38I see the connection now. Right, because to execute that
- 9:40single request, the agent is spawning multiple background
- 9:44threads simultaneously. One thread is pinging airline AP
- 9:47is and others querying hotel databases. 1/3 is cross
- 9:50referencing your meeting schedule.
- 9:52If each of those sub agents takes 5 seconds to query, reason
- 9:56and respond, the entire orchestration layer just
- 9:59collapses into a huge bottleneck.
- 10:00Right, you'd be waiting forever. Exactly.
- 10:03Real time agentic behavior requires near 0 latency at the
- 10:07inference level. And the crazy thing is, Pichai
- 10:09notes, he is already living in this environment.
- 10:12He relies on an internal system occasionally referred to as
- 10:15antigravity or jet ski. Jet ski, yeah, yeah.
- 10:19Where a cluster of these AI agents collaborate to manage his
- 10:23complex daily executive workflows.
- 10:26It's a complete shift from an information retrieval paradigm
- 10:30to a task execution paradigm. It really is.
- 10:34But, you know, running millions of autonomous agents executing
- 10:37parallel API calls for billions of users requires just a
- 10:41staggering amount of compute. Which brings us right back to
- 10:44the $180 billion capital expenditure.
- 10:47Exactly. And here's where it gets really
- 10:49interesting. Because despite spending all
- 10:51that money, there's still a vocal subset of the tech
- 10:54community that claims Google isn't fanatical enough about AGI
- 10:58artificial general intelligence. Yeah, you hear that critique a
- 11:01lot. It's a critique that feels
- 11:02increasingly absurd and detached from reality when you hear
- 11:06Pichai list the internal milestones he's witnessed over
- 11:08the last decade. I mean he has had a front row
- 11:10seat to the fundamental breakthroughs in the field.
- 11:13Literally in the room. He talks about his AGI moments
- 11:15like being there in 2012 when Jeff neural networks
- 11:18spontaneously learn to recognize a cat just by processing raw
- 11:21YouTube data. A huge moment.
- 11:23Then lynching spatial awareness develop in DARPA robotics in
- 11:262014, watching Demesis Savas's early deep mind models execute
- 11:30strategies that literally mimic human imagination.
- 11:34And now today he's observing AI right test and deploy complex
- 11:38code completely autonomously without any human integrated
- 11:43development environment. Watching an architecture write
- 11:46and debug its own software without an IDE is that's a
- 11:50definitive AGI milestone. But this is where the interview
- 11:53pivots from software theory to a brutal reality check.
- 11:57The physical reality. Right.
- 11:58We're sitting here in early 2026, and the defining feature
- 12:01of the industry right now isn't algorithmic innovation or
- 12:04software, it is physical supply constraints.
- 12:07Right, the laws of physics are actively bottlenecking the
- 12:09exponential curve. Exactly.
- 12:11But Chai states plainly that even if you wanted to allocate
- 12:14$400 billion this year, the global supply chain couldn't
- 12:17physically absorb it. Wow, even if they have the cash
- 12:20ready to go? Yep.
- 12:21He breaks down 4 specific hard bottlenecks choking the hardware
- 12:25ecosystem. Let's examine the mechanics of
- 12:27those because they are fascinating.
- 12:28The first is wafer production at the foundries right.
- 12:31Right, we are bumming up against the physical limits of EUV,
- 12:34Extreme ultraviolet lithography and advanced packaging
- 12:37techniques. You cannot simply construct a
- 12:40new TSMC foundry overnight to meet demand.
- 12:43Yeah, it takes what, four to five years to build the clean
- 12:45rooms? At least four to five years to
- 12:48install the optics and calibrate the manufacturing process just
- 12:51to yield usable logic chips. And even if you secure the logic
- 12:55chips, you hit the second bottleneck memory.
- 12:58The interview highlights severe supply shortages and high
- 13:01bandwidth memory, or HBM. Which is absolutely essential.
- 13:04Right, because you need it for feeding data into the processors
- 13:08fast enough to prevent idle time.
- 13:10And HBM manufacturing is incredibly complex.
- 13:12It requires stacking multiple memory chips vertically and
- 13:15connecting them with microscopic vias that run directly through
- 13:18the silicon. Yield rates are terrible.
- 13:20Right. Yeah, notoriously difficult to
- 13:22optimize. The percentage of chips that
- 13:25actually work without defects is low.
- 13:28Demand for HBM is vastly outstripping the foundry's
- 13:31ability to yield it, and that shortage is expected to persist
- 13:35well into 2027. OK.
- 13:37So that's two. The third bottleneck shifts from
- 13:40silicon to physical infrastructure, power and energy
- 13:44constraints. This is a huge one.
- 13:46It's not just an issue of generating enough electricity,
- 13:49it's the transmission. There is a massive global
- 13:52backlog for industrial electrical Transformers.
- 13:54Yeah, you can't get them anywhere.
- 13:55Combine that with the years of regulatory permitting and delays
- 13:59required to hook a GW scale data center into a local grid, and
- 14:03the timeline just stretches out significantly.
- 14:06Which feeds directly into the 4th constraint, societal
- 14:09pushback. We are seeing actual legislative
- 14:12bands and moratoriums on new data centers in places like
- 14:15Ireland and parts of the United States.
- 14:17Yeah, local municipalities are pushing back hard against the
- 14:20strain drain on their power grids and the sheer volume of
- 14:22water required for the cooling. Tower, They just don't want them
- 14:25in their backyards. But you know, I'd argue there's
- 14:28a silver lining here. These hard physical constraints
- 14:31are actually forcing innovation because they can't just buy
- 14:34their way out with infinite compute.
- 14:36They are being pushed into algorithmic elegance.
- 14:38Absolutely. The goal is a 30X increase in
- 14:41energy efficiency. Right, pushing models to be
- 14:44lighter and open source. Look at a model like Google's
- 14:47Gemma 4. That is a perfect example of
- 14:49constraint breeding innovation. Gemma 4 is the top open source
- 14:53model outside of China right now.
- 14:55It's built on the massive Gemini 3 architecture, but it utilizes
- 14:59advanced distillation techniques to compress the reasoning
- 15:02capabilities into a much smaller footprint.
- 15:05You get a lightweight model that can run locally, dramatically
- 15:08reducing the reliance on those bottlenecked centralized data
- 15:11centers. So, OK, if the infrastructure of
- 15:13Earth is tapped out, if we are running out of physical space,
- 15:16power, capacity and community patients for data centers, where
- 15:20does a company with an unlimited budget look next future, the
- 15:23future? And this brings up their 10 to
- 15:2520 year moon shot projects, which I'm so excited to talk
- 15:28about. Yeah, it's essential to
- 15:30understand Google's capital allocation philosophy here.
- 15:32They don't look at short term ROI for these, they look at long
- 15:36term option value and technical milestones.
- 15:39They're placing bets on technologies that can bypass
- 15:42these physical limits entirely. And one of those bets is
- 15:44literally taking the data centers off the planet.
- 15:48Space data centers. It sounds like pure science
- 15:50fiction. It does, but when you look at
- 15:52the physics, the logic really holds up.
- 15:55Space provides uninterrupted solar energy and an ambient
- 15:59environment that is near absolute zero.
- 16:01It literally immediately solves the power generation and thermal
- 16:04cooling crises. Right.
- 16:06The primary engineering hurdle there isn't power, it's data
- 16:09transmission. Like how do you move petabytes
- 16:12of data from orbit to ground stations without crippling
- 16:15latency? Right, you can't just run an
- 16:17Ethernet cable. No, you can't.
- 16:18It requires a massive leap in optical communication utilizing
- 16:22laser links between satellite constellations rather than
- 16:25traditional radio frequencies. It is still in the very early
- 16:28feasibility stages, but the underlying physics actually
- 16:31support the architecture. They are also looking at compute
- 16:34paradigms that don't rely on traditional silicon logic at
- 16:37all. Quantum computing, right?
- 16:40Pichai outlines their specific aim here, simulating incredibly
- 16:44complex systems like molecular biology and global weather
- 16:47patterns. But the metric they're tracking
- 16:50isn't commercial viability yet. It is breaking the error
- 16:53correcting Quibit threshold. And understand why that matters.
- 16:57You really have to look at how unstable quibits are.
- 17:00A physical quivet loses its quantum state it deco.
- 17:03Here's that the slightest temperature fluctuation or
- 17:07electromagnetic interference. It's super fragile.
- 17:10Incredibly fragile. So the engineering milestone
- 17:13isn't just cramming more physical quibits onto a
- 17:15processor. It is grouping hundreds of
- 17:17physical quibbits together to create one stable, logical
- 17:21quibit that can correct its own errors in real time.
- 17:24Oh wow. And once you cross that
- 17:25threshold of stability, the compute capacity scales
- 17:28exponentially, right? Which ties directly into their
- 17:30other bets. Look at Isomorphic Labs, which
- 17:32is using these advanced AI predictive models to completely
- 17:36overhaul pharmaceutical drug discovery.
- 17:38Yeah, that's huge. Or how they're injecting
- 17:40Gemini's spatial reasoning models into Boston Dynamics
- 17:43platforms, allowing robotics to navigate the physical world with
- 17:47just unprecedented contextual awareness and wing drones
- 17:51expanding delivery to 40 million Americans.
- 17:54It's all connected. All of these isolated
- 17:56advancements, the low latency inference, the asynchronous
- 17:59agents, the quantum research, the robotics, they're all
- 18:02converging on a very specific timeline.
- 18:04This is the climax of the deep dive.
- 18:06Yes, Pichai definitively predicting that 2027 will be the
- 18:10enterprise AI singularity. 2027, that is next year.
- 18:15Let's define what that actually means in a practical sense.
- 18:18Right now, AI is largely A localized tool for developers.
- 18:21It helps a code to write faster or marketer draft copy.
- 18:23It operates within a silo. But in Pichai's 2027 prediction,
- 18:27AI transitions from a localized developer tool into a non
- 18:30engineer smart brain. It becomes the overarching
- 18:33connective tissue. Currently, corporate data is
- 18:36heavily siloed. Your supply chain data sits in
- 18:39one ecosystem, HR data in another.
- 18:41Financial projections in 1/3. The Agenic manager will natively
- 18:45collapse all of those silos. It will be capable of Florida
- 18:49State executing complex workflows that span multiple
- 18:52departments simultaneously. And to ensure that integration
- 18:55actually functions Florida State, you have to look at
- 18:57Pichai's personal innovation habit.
- 19:00He relies heavily on the engineering practice of dog
- 19:03fooding. Yeah, he forces himself to use
- 19:05the internal test versions. Right, he doesn't just read
- 19:08reports, he actively uses the unstable alpha builds.
- 19:11He famously mentions spending 30 minutes at the gym just chatting
- 19:15with experimental builds of Gemini Live while stretching.
- 19:18Just actively probing the architecture for conversational
- 19:20latency and logic flaws while working out.
- 19:23Exactly, it is the only way to genuinely understand the
- 19:26friction points when the CEO is testing edge cases during a
- 19:29workout and manually rational TPU chicks like headcount.
- 19:32It underscores the sheer intensity of the transition we
- 19:35are currently. In they are balancing the
- 19:37demands of billions of users, rigorous safety protocols, and
- 19:40brutal physical supply chain limits, all to usher in an era
- 19:44where AI agents seamlessly manage our entire workflows by
- 19:472027. Absolutely.
- 19:50And before we get to our final thought for today, just a quick
- 19:52note for you all to access more advanced researches,
- 19:55applications and our latest published personal intelligent
- 19:58Multi Agent Deep Value Intelligence.
- 20:01Welcome to explore and join the Deep Values dot AI community.
- 20:05Yeah, and to track total market temperature, welcome to search
- 20:08and download the iOS App Buffett Indicator from the Apple App
- 20:12Store. Just a quick disclaimer, all
- 20:14analysis is based on publicly available sources and may be
- 20:17subject to revisions or differing interpretations.
- 20:20All contents from this channel is for educational and research
- 20:23purposes only and does not constitute investment advice.
- 20:26Conduct your own due diligence before making investment
- 20:29decisions. And if you'd like to access
- 20:31wider topics, including life and investing philosophy, please
- 20:33subscribe to our sister channel Deep Values.
- 20:35Don't forget to like this video. Much appreciated.
- 20:38So as you look at your own daily professional environment, you
- 20:42really need to identify the tasks you were doing today.
- 20:45Reconciling data, coordinating logistics, drafting routine
- 20:49analysis, right? Because the shift isn't about
- 20:52learning how to do those tasks slightly faster.
- 20:54It is about preparing to transition from executing the
- 20:57work yourself to managing the agent that will likely handle it
- 21:01entirely asynchronously next year.
- 21:03We are rapidly building this hyper efficient architecture.
- 21:06Everything is being optimized, latency is being driven to 0 and
- 21:10workflows are being Florida State automated.
- 21:12It is a remarkable engineering achievement, but it does leave
- 21:15us with a broader, final, provocative thought regarding
- 21:18this perfectly optimized world we are building.
- 21:20Yeah, this is something I want everyone listening to them all
- 21:23over on their own. If these AI agents become
- 21:25perfectly Florida State capable of automating and orchestrating
- 21:28all our professional and personal workflows without
- 21:30deviation, what happens to the value of human error?
- 21:35That's a great question. Right.
- 21:36So much of human innovation, scientific discovery and
- 21:39creative progress has historically been the direct
- 21:42result of accidental detours, miscommunications and
- 21:45serendipitous mistakes. If we engineer all the friction
- 21:48out of the system, will those serendipitous mistakes that
- 21:51drive human creativity just be lost?
- 21:53Something to ponder as we approach the Singularity.