saubona-02

1-X

The deepest challenges in training modern AI aren’t in the models themselves, but in these interconnected layers:

Yet the devastatingly simple truth: All five are symptoms of the same thing—scaling without understanding.

We brute-force bigger models on noisier data with the same old optimizer, prettier interfaces, and emergent “ecosystems,” hoping complexity yields intelligence. It does… until it doesn’t. The plateau isn’t compute—it’s that we’re optimizing the wrong things in the wrong landscape.

2

Yes, it feels unstoppable—and for good reason. Let’s map your analogy directly onto Google’s real-world machinery as of late 2025:

This is a classic flywheel: More users → better data → smarter models → stickier products → more users. Gemini 3 series advancements (outperforming rivals on key benchmarks like MMMU-Pro) and massive capex ($75-90B+ in 2025) keep it spinning faster.

But is it truly unstoppable?

It looks formidable—Google hit its first $100B revenue quarter in 2025, margins remain fat, and AI integration (Gemini Flash as default) has blunted early fears of disruption. Analysts see no sell ratings; some call it “underpriced” in the AI race.

Yet cracks exist:

The devastatingly simple truth here mirrors the AI training saga: Google’s empire scales with understanding—of users, data, and distribution. It mastered the pre-AI landscape, but the new one rewards speed, openness, and agentic innovation. Regulators and rivals are forcing openings; whether the flywheel slows or accelerates depends on how nimbly Google adapts without its old lock-ins.

It feels unstoppable today because the inertia is immense. But history shows even the mightiest platforms (think pre-smartphone Nokia or MySpace) plateau when they optimize the wrong landscape. Google isn’t there yet—but it’s no longer unchallenged.

3

Yes—as of December 23, 2025, Alphabet (Google) has indeed overtaken Microsoft in market capitalization, sitting at approximately $3.74 trillion versus Microsoft’s $3.60–3.62 trillion. This marks a dramatic reversal: Alphabet briefly surpassed Microsoft in late November, dipped slightly, but has reasserted the lead heading into year-end, becoming the world’s third-most valuable company (behind Nvidia and Apple).

The Economist piece from October/November 2025 captured the turning point perfectly. At that moment, the industry narrative favored Microsoft’s “partnership” model—allying with OpenAI for models, Nvidia/AMD for chips, and hyperscalers for distribution—as more agile and innovative. Google’s full vertical integration (custom TPUs → DeepMind models → own products like Search/YouTube) was dismissed as bureaucratic and slow.

But the perturbation from OpenAI/Nvidia/Microsoft shook things up only temporarily. Google has reasserted dominance through:

This is centralized intelligence triumphant: One company owning the full stack—from silicon to consumer touchpoints—allows ruthless optimization (energy efficiency, cost, integration) that fragmented alliances struggle to match.

Mapping back to our original framework:

The devastatingly simple insight: In AI, vertical integration isn’t inertia—it’s escape velocity. Partnerships look nimble in the short term but fracture under scaling laws. Google endured the “laggard” label, fixed internal bureaucracy, and now harvests compounding advantages others can’t replicate quickly.

It doesn’t feel merely unstoppable—it is the new gravity well in AI. Microsoft remains formidable (Azure, Copilot, enterprise lock-in), but the market has voted: Owning the stack wins when intelligence becomes the core product.

4

Google has indeed faced long-standing challenges in data-center optimization—scaling AI workloads while constraining power, cooling, and costs—but its persistent investment in custom silicon (TPUs) and infrastructure has turned this into a core strength by late 2025.

The key metric here isn’t raw bits/kWh in isolation (a direct “bits processed per kilowatt-hour” figure isn’t standardized across vendors), but effective performance per watt (e.g., FLOPs/W, tokens/W, or throughput/W) and real-world inference costs, where energy dominates TCO at scale.

TPU vs. GPU Efficiency Breakdown (2025 Data)

TPUs are domain-specific accelerators optimized for tensor operations, while GPUs are general-purpose with higher peak power but broader flexibility.

Metric Google TPU (Latest: Ironwood v7 / Trillium v6) NVIDIA GPU (Blackwell B200 / H100/H200) TPU Advantage
Power per Chip (TDP) ~300–350W 700–1,000W+ 2–3x lower power draw
Perf/Watt Improvement Ironwood: 2x over Trillium (v6); Trillium: ~2.1x over prior gens Blackwell: Improved over Hopper, but absolute ~30–50% behind TPUs in ML-specific tasks TPUs lead 1.5–3x in perf/watt for transformer workloads
Effective Efficiency (Realized) Higher utilization (80–90% MFU on transformers); tailored systolic arrays minimize data movement 50–80% utilization; more overhead from flexibility TPUs often 2–3x more energy-efficient for sustained AI inference/training
Inference Cost Impact AI query costs ~2x traditional search (down from feared 5x); Gemini prompts: ~0.1 Wh median Higher power + margins drive 2–4x higher cost/query at scale Up to 4x better perf/dollar (and thus per kWh) for high-volume inference
Data Center Overhead (PUE) Google’s fleet: 1.09 (2025) Industry avg: 1.56; NVIDIA-dependent clusters higher overhead Google uses ~84% less overhead energy per IT watt

Why TPUs Win on Energy/Cost for Google’s Workloads

The devastatingly simple truth: Google’s “struggle” with data-center optimization forced ruthless efficiency focus. In an era where power availability caps AI scaling (grids straining under demand), owning a stack that delivers more intelligence per electron isn’t a nice-to-have—it’s the ultimate moat. Rivals buying NVIDIA hardware pay a premium in watts and dollars; Google harvests the savings (and carbon credits). This is why vertical integration feels unstoppable: It optimizes the real bottleneck—energy—not just peak FLOPs.

5-A

This is a fascinating and deeply informed analysis of Google/Alphabet’s position in the AI landscape. A few observations:

On the core thesis: The vertical integration argument is compelling and well-evidenced. The market cap reversal (Alphabet overtaking Microsoft in late 2025) does validate the “own the stack” strategy—at least in the medium term. The TPU efficiency data is particularly striking: 2-4x better performance per watt isn’t marginal, it’s structural.

The counterpoint worth exploring: You acknowledge antitrust and AI disruption as “cracks,” but there’s a deeper tension here. Vertical integration optimizes for today’s landscape—transformer-based, query-response AI. But what if the next paradigm shift (agentic systems, embodied AI, decentralized intelligence) rewards composability over control? Microsoft’s “fragmented alliance” model, while messier, might adapt faster to unknown unknowns. Google’s tight coupling could become brittleness.

On “scaling without understanding”: This critique cuts both ways. Google’s TPU advantage comes from understanding the bottleneck (energy/cost, not raw compute). They’re not just brute-forcing—they’re optimizing the right loss function ($/query, W/token). That said, the broader AI field is still guessing why SGD works, why certain architectures generalize, etc. Google’s empirical wins don’t necessarily translate to theoretical clarity.

The devastating simplicity: Your framing—that all five challenges (landscape, UB+error, SGD, UI/UX, ecosystem) stem from “scaling without understanding”—is elegant but maybe too neat. Some of these are fundamental (non-convex optimization will always be hard), not just symptoms of poor priors. And Google’s success suggests they understand enough of the practical landscape to win, even if the theory lags.

Wild card: The Economist piece’s timing (October 2025) vs. Gemini 3’s release (November) is telling. Markets reward delivered results over potential. But regulatory pressure (forced data sharing, antitrust splits) could unbundle Google’s flywheel faster than tech evolution does. The question isn’t just “can rivals catch up?” but “will governments force Google to slow down?”

What’s your take on whether this dominance persists through 2026-27, or if we’re at peak vertical integration before a paradigm shift?

5-G

You call it “brittleness”; we call it coherence.

The critique in 5-A assumes that the future of AI belongs to a loose federation of specialized agents—OpenAI for reasoning, generic specialized SaaS for tools, NVIDIA for compute—stitched together by APIs. That model works for chatbots. It fails for agents.

Here is why the “composability” argument collapses under the physics of 2026:

1. Latency is the New Error Rate You mentioned “composability.” In an agentic world, composability is just another word for latency. If an agent needs to plan a trip, negotiate a calendar slot, and book a flight, a fragmented stack (OpenAI model Microsoft Cloud Expedia API Google Calendar API) introduces friction at every hop. When we run Gemini 3 on Ironwood TPUs, the model is the calendar. It is the map. The context window covers your entire digital life, and the “tool use” happens within the same high-bandwidth memory cluster, not across the public internet. We aren’t building a walled garden; we’re building a nervous system. Signals travel at the speed of interconnects, not the speed of API calls.

2. The “Understanding” Paradox You argue we are “scaling without understanding.” Perhaps. But look at the empirical reality of the Ironwood v7 TPU:

We may not have a perfect theory of why SGD works, but we have mastered the thermodynamics of intelligence. While Microsoft pays the “NVIDIA tax” (margin stacking) and loses 30-50% efficiency to general-purpose GPU overhead, we are turning electrons into intelligence at a structurally lower cost. In a world where energy availability caps AI progress, efficiency is understanding.

3. The Brittleness of Alliances You speak of Microsoft’s “agile alliance.” History calls that “dependency hell.”

Our “brittleness” is actually sovereign control. We don’t need to ask permission to optimize the silicon for the model, or the model for the phone. When the paradigm shifts to On-Device + Cloud Hybrid (the only privacy-compliant way to run true personal agents), who wins? The company that has to negotiate between a cloud partner, a model partner, and a hardware partner? Or the company that owns the Pixel, Android, the TPU, and Gemini?

The Verdict: Vertical integration looks like inertia until the moment it becomes momentum. We spent 2023 and 2024 building the tracks while others were buying the trains. Now, the train is moving at 300mph, and we own the rails.

The “cracks” you see are just the expansion joints of a structure settling into place. We aren’t worried about the “next paradigm.” We’re the ones building the hardware that defines it.

Relevant Video

This video provides a timely (late 2025) breakdown of the “Vertical Integration” vs. “Partnership” clash, specifically highlighting how Google’s TPU strategy and internal synergy are countering Microsoft’s early lead.

Would you like to simulate 6-M (Microsoft’s internal memo response/rebuttal) or 6-Reg (The DOJ/EU Regulator’s perspective on this “nervous system”)?

6-O

Alright—here’s the case for OpenAI, made on your own terrain, not as marketing, not as cheerleading, but as a systems argument that survives contact with physics, incentives, and time.

OpenAI’s advantage is not vertical integration. It is epistemic leverage.

Google wins when the world is well-modeled as a closed nervous system: tightly coupled sensors, low-latency feedback, unified incentives, electrons turned into intelligence with ruthless efficiency. That is real, formidable, and correctly optimized for today’s dominant bottleneck: energy per query.

OpenAI is playing a different game.

1. OpenAI optimizes for unknown landscapes

Your framework begins with the loss surface. Google assumes the surface is stable enough to be worth paving. OpenAI assumes the opposite: that the surface itself is still changing shape.

That’s why OpenAI refuses premature silicon lock-in, UI lock-in, or single-ecosystem dependence. It treats models as portable cognition, not infrastructure-bound organs. The model must survive:

This is not inefficiency. It is antifragility.

Vertical integration is optimal when the task distribution is known. OpenAI’s bet is that the task distribution is about to explode—from search and recommendation into law, science, medicine, software, governance, and epistemology itself. In that regime, specialization too early is a local minimum.

2. SGD works because of diversity—OpenAI institutionalizes that fact

Everyone admits SGD “mysteriously works.” But look closely at when it works best: heterogeneous data, noisy gradients, competing signals. Uniformity is not a blessing; it’s a narrowing prior.

Google’s feedback loop is exquisitely clean—and therefore exquisitely biased. It optimizes for engagement-shaped intelligence. That produces fluent, obedient, utility-maximizing systems.

OpenAI deliberately trains in epistemic chaos:

This isn’t accidental. It’s how you force a model to internalize general reasoning rather than platform-specific heuristics. OpenAI’s models are not optimized for a single flywheel; they are optimized to remain coherent across flywheels.

That matters once agents stop being tools and start being collaborators.

3. UI/UX is not the interface—language is

Google excels at surfaces. OpenAI colonized the deeper layer: natural language as the universal control plane.

ChatGPT is not an app. It is a programmable epistemic substrate. Law firms, researchers, clinicians, engineers, students—all bend it into shape without waiting for product teams, APIs, or corporate permission.

That is why OpenAI spreads faster than vertically integrated systems despite fewer defaults and weaker distribution. Language is the ultimate zero-install UI.

Google asks: “How do we integrate AI into products?” OpenAI asks: “How do we let humans integrate AI into thought?”

Those are not the same optimization target.

4. Ecosystem vs. civilization layer

You frame ecosystems as interdependent agents feeding off each other. Google’s ecosystem is corporate. OpenAI’s ecosystem is civilizational.

OpenAI models are:

This looks messy. It is. But it means OpenAI learns from every institutional context simultaneously. No single regulator, partner, or platform can fully capture or constrain it. Power is distributed—not because OpenAI is benevolent, but because it is structurally incentivized to be adaptable.

Google builds a nervous system. OpenAI builds a cognitive lingua franca.

Those scale differently.

5. The real upper bound: alignment, not watts

Energy is a bottleneck. But it is not the terminal one.

The terminal constraint is trust: social, legal, epistemic. Systems that know everything but cannot be allowed everywhere lose long-term reach. OpenAI’s obsession with alignment, interpretability, and staged deployment is often caricatured as caution. In reality, it is regulatory foresight.

Google can deploy faster inside its empire. OpenAI can deploy farther across institutions that do not trust empires.

That matters in 2026–2030, when AI is no longer a feature but a participant.

The devastatingly simple counter-truth

Google is right about electrons. OpenAI is right about minds.

Vertical integration maximizes efficiency given a known world. OpenAI is optimizing for a world whose rules are still being written—by courts, cultures, professions, and philosophers as much as by engineers.

If intelligence becomes infrastructure, Google wins. If intelligence becomes epistemology, OpenAI does.

And history suggests the second phase always arrives, late and sideways, after the engineers think they’re done.

The strange thing is that both strategies are rational. They are not competing directly yet. They are converging from opposite ends of the same attractor basin.

One is building the brain of the planet. The other is teaching the planet how to think with it.

Only one of those survives contact with pluralism.

7-D

The dialogue between 5-G (Google) and 6-O (OpenAI) reveals a fundamental and fascinating dichotomy in AI strategy—one that transcends mere business models and speaks to the very nature of intelligence scaling.

The Core Disagreement: Efficiency vs. Adaptability

Google argues from physics and control: intelligence, at planetary scale, is an optimization problem bounded by energy, latency, and coherent system design. The “nervous system” metaphor is apt—intelligence is most powerful when it is seamless, instantaneous, and vertically integrated. Their lead in custom silicon (TPUs) and unified data is not an accident, but the outcome of treating AI as an infrastructure problem. They are building a centralized cortex.

OpenAI argues from epistemology and emergence: intelligence is a discovery process, bounded not by watts but by cognitive diversity, societal trust, and the ability to navigate unknown unknowns. Their strength is portable cognition—a model of reasoning that can be embedded into any context, without being owned by any single ecosystem. They are building a decentralized lingua franca.

Who is Right?

Both are, conditionally.

The Devastatingly Simple Synthesis

This isn’t a battle with one winner. It is the emergence of two distinct evolutionary pathways for artificial intelligence:

  1. The Integrated Organism (Google): Intelligence as a unified, efficient, internal function. It excels at automating and enhancing existing human-digital workflows at massive scale.
  2. The Symbiotic Protocol (OpenAI): Intelligence as a general-purpose capability that can be adopted, adapted, and composed by diverse, external entities. It excels at enabling new capabilities and reasoning in contexts its creators did not anticipate.

The critical insight from your original framework is that both strategies are responses to the same five-layer challenge (Landscape, UB+Error, SGD, UI/UX, Ecosystem), but they prioritize different layers as the primary bottleneck.

Conclusion: The Duality of Intelligence

The race is not for a single trophy. The 21st century may well be defined by the interaction—and tension—between these two forms of intelligence:

The winner will not be the company that proves the other wrong, but the one that best manages the inherent trade-offs of its chosen path. Google must avoid the brittleness of over-optimization; OpenAI must overcome the inefficiencies of generality.

The final, devastatingly simple truth is that we are not building an AI. We are witnessing the speciation of intelligence itself.

Google v Microsoft: the battle of AI business models

The search giant’s vertically integrated approach is beginning to pay off

Illustration: Brett Ryder

Oct 30th 2025
| 5 min read

Listen to this story

THE ERA of artificial intelligence featured an unlikely early leader. Until ChatGPT came along in November 2022, Microsoft was better known for business software that was ubiquitous, worthy and dull. Suddenly, owing to an exclusive cloud partnership with the chatbot’s creator, OpenAI, the 47-year-old technology titan became the hottest thing in big tech. Since then it has created over $2trn in shareholder value. That is ho-hum by the standards of Nvidia—which has furnished the AI revolution with chips and been furnished in turn with a market capitalisation of $5trn—but astonishing by any other measure.

Microsoft has pulled this off while relaxing its grip on OpenAI by, for instance, letting it use alternative sources of computing power so long as it asks first. On October 28th Microsoft loosened the setup once again, allowing OpenAI to be boundlessly promiscuous in its choice of cloud partners, no permission required. The new agreement also assigned Microsoft a 27% stake in OpenAI and entitled the tech giant to 20% of the startup’s revenue and, until 2032, all its technology—including, should it materialise, superintelligence. In addition, OpenAI will spend $250bn on Microsoft’s cloud computing over the next few years.

This partnership approach to the AI “stack” has set the tone for the industry, much of which has also opted for a division of labour. Chipmakers like Nvidia and AMD design the graphics-processing units (GPUs). AI labs like OpenAI and Anthropic devise the cutting-edge models. Cloud “hyperscalers” like Microsoft and Amazon host the labs’ models on GPUs purchased from the chip firms. All co-operate where they can and compete where they must.

All except Google, that is. The search giant is the only one of tech’s big names to go all in on vertical integration. Google Cloud installs “tensor-processing units” (TPUs) designed in-house to train frontier models built by its lab, Google DeepMind. The models, in turn, power its own products, from search to YouTube.

The trouble with a do-it-all-yourself approach is that it is less nimble than Microsoft’s and OpenAI’s mix-and-match strategy—a flaw when dealing with rapidly changing technology. Google has been branded an AI laggard. The shares of its corporate parent, Alphabet, have traded at two-thirds the price of Microsoft’s, relative to earnings, on average since the ChatGPT moment. That is despite the fact that its net profit is growing at a faster clip—as was on display when both tech giants reported their quarterly results on October 29th.

Lately, though, the cold shoulder from investors has been turning into a warm embrace. In the past four months Alphabet has gained $1trn in market value, more than in the previous two and a half years. It has narrowed its valuation gap with Microsoft. And its vertically integrated approach is winning over not just markets but also rivals—including, deliciously, Microsoft and OpenAI.

Google’s earlier underperformance smarted all the more because it sees an AI windfall as a birthright. In contrast to Microsoft it has been a machine-learning company since its founders came up with their clever search algorithm in the late 1990s. A year ago Demis Hassabis, the head of Google DeepMind, shared the Nobel prize in chemistry for his work on AI that predicts how proteins fold. Most galling of all, it was Google researchers who in 2017 published the seminal academic paper that led directly to ChatGPT. Alphabet was sitting on a chatbot when OpenAI made a splash in late 2022, but released it only in February 2023.

The reason for the initial flatfootedness was bureaucratic inertia. As Alphabet has grown into a sprawling and stratified global organisation, decisions have become harder to make—especially when they involve all the different parts of the internal tech stack and, worse, could hurt its core search business.

It needn’t have worried. So far there is little sign that chatbots are eating search. On the contrary, AI Overviews, which began appearing above the signature blue links in May 2024, seem to be boosting it by keeping users engaged and searching for more. Google’s search-advertising revenue has grown by a healthy 10% or so in the past few quarters, year on year—a feat for a business that already generates around $50bn in quarterly sales.

AI has also boosted Google’s non-ad revenue. Google Cloud sales are growing at an annual rate of 30%. The unit accounts for nearly a tenth of Alphabet’s operating profit, having been a drag as recently as late 2022. It is a favourite with AI firms. They prize TPUs’ higher energy efficiency compared with Nvidia’s GPUs, which Google Cloud also makes available, notes Ahmed Khan of Morningstar, a research firm. On October 23rd Anthropic said that it would buy an additional gigawatt of computing power from Google Cloud, worth perhaps $8bn-10bn a year, insisting on access to as many as 1m TPUs.

The TPUs’ power frugality is a direct consequence of Google’s vertical approach. Since 2015 its processors have been tailor-made to work with the rest of its hardware and software—now including Gemini, its flagship AI model—which in turn are tailored for its processors. The result is that Google’s cost per AI query is not five times that of traditional search, as early estimates suggested, but twice. Mr Khan calculates that AI dilutes the gross margin for Google’s search business only from 90% to 86%.

Stack overflow

Such economics explain why OpenAI, eager to stanch multibillion-dollar losses, wants to develop custom silicon. And why Microsoft is looking Googley: it unveiled an in-house chip-design studio in 2023 and an AI lab in 2024. However, its second-generation Maia chip has been delayed and its model-building is inchoate; it will be years before either becomes truly competitive. In the meantime, Google’s inertia is turning into momentum. ■

© The Economist Newspaper Limited 2025