What Remains After the Showmanship Fades: Google’s 10-Year Counteroffensive by Design

ChatGPT, Claude, Gemini, and even Cursor and various APIs.
I subscribe to virtually every major AI paid service that exists.

Driven by a desire not to miss out on new technology, I spend hundreds of dollars monthly,
but thanks to this, I’m among the most sensitive to detect subtle changes in each model.

Recently, a seismic shift occurred in my ‘AI usage share’.

Until last year, ChatGPT was undoubtedly the center of my work,
with Claude backing it up as a coding partner.

At that time, Gemini was merely a ‘benchmark testing’ toy
that I occasionally fired up to keep track of the latest tech trends.

But now, the central axis of my workflow has completely shifted.

Gemini (50%) - Claude (40%) - ChatGPT (10%)

This change didn’t happen overnight as a sudden change of heart.

There was a clear ‘experiential shift’ in 3 phases starting from January 2025,
and looking back on this, I could see the grand picture Google had been preparing for 10 years.

1. The Winner’s Formula: Vertical Integration

History Repeats Itself

Companies that have dominated massive industries share a common trait.

They identified elements that would be difficult to establish with the current ecosystem alone, then built all of those elements themselves and connected them.

Apple’s Choice (Apple Silicon)

The problem Apple saw: Intel chips can’t deliver the user experience I want to create.

• Strategy:
  • Custom chip design (M1, M2, M3)
  • macOS optimization
  • Complete hardware design
  • Integration with their already exceptional UI/UX
• Result:
  • ‘Performance per watt’ that Intel chip-using competitors can’t match
  • An experience where chip-OS-design operates like a single organism
  • Apple has cornered the market on integrated mobile and desktop experiences at the chip level.

Tesla’s Choice (Full Stack)

The problem Tesla saw: The existing combination of parts can’t sustain the electric vehicle industry.

• Strategy:
  • Camera-centric autonomous driving: Cameras only instead of expensive LiDAR
  • Real-time driving data pipeline from millions of vehicles
  • Custom AI chip design (FSD Computer)
  • Custom battery development (4680 cells)
  • Giga Press: Manufacturing cost innovation through single-piece casting
  • Supercharger Network: Direct operation of charging infrastructure
• Result:
  • Autonomous driving equal to or better than LiDAR at 1/100th the sensor cost
  • Overwhelming manufacturing cost reduction
  • Data scale and charging experience competitors can’t match
  • Redefinition of the entire electric vehicle industry

Google’s Choice (AI Infrastructure)

The problem Google saw: NVIDIA GPUs can’t sustain the economics of the AI industry.

• Strategy:
  • Custom chip design (TPU v1~v6, 10 years)
  • Custom datacenter operation
  • Compiler (XLA) optimization
  • Creation of Transformer architecture
  • Monopoly on YouTube/Search/Maps data
• Result:
  • Overwhelming cost efficiency compared to NVIDIA dependency
  • API pricing competitiveness
  • Securing business viability (BM) of AI services

The commonality is clear.

They didn’t simply create ‘better products.’

They identified all elements necessary for the industry to exist, and directly controlled everything from the lowest level of infrastructure to user experience.

This appears to be an inefficient investment in the short term,
but once it crosses a critical threshold, it becomes an ‘unreplicable moat’.

2. 10 Years of Accumulation: ‘Depth of Knowledge’ Beyond Hardware

We commonly find Google’s strength
only in physical infrastructure like ‘TPU (custom chips)’ and ‘datacenters.’

But what’s truly formidable isn’t the hardware itself.

It’s the ‘depth of optimization knowledge’ accumulated
by running that hardware for 10 years.

‘Connected Knowledge’ Emerging from Full Stack

For competitors, GPU manufacturers (NVIDIA), cloud operators (MS/AWS),
and model developers (OpenAI) are separated.

They communicate only through APIs, ignorant of each other’s black boxes.

In contrast, all layers are open at Google.

• Chip designers know Transformer computation characteristics and design circuits accordingly
• Compiler engineers know the chip’s physical characteristics and optimize code
• Datacenter operators know model heat patterns and run cooling systems accordingly

This ‘Cross-domain Knowledge’ cannot be bought with money.

Google’s formidability comes from here.
The optimization capability that breaks down boundaries between hardware and software.

This is the ‘invisible depth’ that competitors cannot imitate.

3. The Three Phases of Experience: Value, Viability, and Powerful Performance

My Gemini usage experience
was like experiencing the moment this ‘accumulated time’ crossed a threshold and exploded.

Phase 1: Discovery of Value (January 2025 - Gemini 2.0 Pro)

When Gemini 2.0 Pro launched early this year,
my perception changed for the first time.

“Huh? This is more useful than I thought?”

At significantly lower cost compared to other models,
it processed quite long contexts.

Even at this point, it wasn’t so much a main model,
but valuable as a ‘sub option’ for saving costs.

Phase 2: Business Conviction (June 2025 - Gemini 2.5 Pro & Flash)

The decisive moment for me as a developer and entrepreneur
was June, with the release of Gemini 2.5 Pro and Flash.

When developing AI services, the biggest headache is ‘cost ratio’.
If the cost per token is high, the business model (BM) doesn’t work.

The performance and pricing policy Google showed at this time was shocking.

“At this price with this performance, we can launch a service with pretty good quality and still have margin.”

This was the point when I deeply experienced Google’s infrastructure efficiency
from the perspective of a provider (Developer), not just a user.

Phase 3: Performance Reversal (Present - Gemini 3.0 Pro)

And now, with Gemini 3.0 Pro,
the paradigm has shifted.

I no longer consider cost ratio or value. I just use it because it’s the smartest.

Complex reasoning, understanding of ultra-long contexts through massive context windows, grasping subtle nuances.
It shows powerful absolute performance that wins even when price is removed from the equation.

The ‘value’ model of January,
through the ‘business tool’ of June,
has risen in November to become my ‘flagship workhorse’.

4. Current Workflow Distribution (50 : 40 : 10)

My current AI workflow
is probably the result of Google’s ‘buildup’ finally converting into personal experience.

1. Gemini (50%) - Main Engine

Role: Research, planning, data analysis

Reason: The smartest, yet fastest and most persistent.
3.0 Pro is a true partner that helps extend my thinking.

2. Claude (40%) - Coding Specialist, Some Writing

Role: Backend of Cursor editor

Reason: Coding is still a domain of ‘precision’ and ‘rigor’.
If Google has advanced in macroscopic reasoning ability, Claude still shows ‘craftsmanship’ at the code block level.

But I can’t guarantee how long this 40% share will be maintained.
Will Google devour even coding with overwhelming computing power, or will Claude defend the specialized ‘craftsman’ domain? This is next year’s point of interest.

3. ChatGPT (10%) - Daily Habit

Role: Light search, small talk

Reason: It’s inertia.
Plus the friendliness and familiarity of the mobile UI.
The mobile app experience seems to be the best designed.
One more thing to add is that the default tone has a strangely friendly feel.

5. Another Moat: The ‘Gravity’ That Attracts Talent

Google’s ‘full stack’ strategy
didn’t just solve problems of money and performance.

Paradoxically, it’s becoming the gravity
pulling back the ‘talent’ that once left Google.

“If Everyone Offers Similar Compensation, What Determines the Choice?”

All big tech companies offer
competitive compensation packages to top AI researchers.

Differentiation through monetary compensation alone is no longer easy.

Yet in recent years, an interesting pattern has been observed.

Noam Shazeer returned to Google from Character.AI,
and the founders of Codeium, which created Windsurf, an agentic coding IDE like Cursor, also joined Google DeepMind.

There’s an increasing trend of startup founders integrating their companies into big tech
and returning to being researchers themselves.

The reason is clear.

“Computing Barrier: An Uncrossable River”

As AI models have grown massive,
a ‘physical gap’ has emerged between startups and big tech.

The scale of infrastructure investment by hyperscalers
is measured in hundreds of billions of dollars.

No matter how much funding a startup receives,
with a structure of renting GPUs from external clouds,
it’s hard to match the ‘experimental speed’ of Google with its custom chips (TPU).

Ideas for research overflow,
but the ‘tools for experimentation’ to prove them are lacking.

Ultimately ‘Infrastructure’ Guarantees Research Freedom

For top-level AI architects,
true welfare isn’t just salary.

It’s ‘sufficient computing resources’
to freely and immediately prove my hypotheses.

When engineers at other companies worry about GPU allocation and adjust experiment scale,
Google’s engineers conduct experiments using large-scale TPU clusters.

An environment where you can voice opinions on chip design to match your model,
and modify the compiler (XLA) for optimization.

This freedom to tinker from the ground up
without boundaries between hardware and software
is the ‘most attractive playground’ for engineers.

This isn’t a romantic homecoming.
It’s a ‘realistic choice’ to complete one’s research.

This infrastructure gap that Google built over 10 years
now provides ‘research freedom’ that can’t be easily replicated even with money,
becoming a powerful factor in retaining talent.

6. The Relationship Between Marketing and Technology: Correcting Misconceptions

Reading this far, you might wonder:

“So is marketing meaningless?”

Absolutely not.

This article doesn’t devalue marketing.
Marketing is an essential medium that delivers innovation to the world.

The Essential Value of Marketing

Without marketing, innovation doesn’t reach the world.

No matter how excellent the technology, if people don’t know about it, it’s as if it doesn’t exist.
Knocking on the market’s door, moving people’s hearts, persuading for change.
This is as important an ability as technology itself.

This is exactly what OpenAI demonstrated with ChatGPT.

They made AI an ‘experienceable reality’ for the public.
For people who said “I don’t know what AI is,”
they made it possible to say “try it once.”

This is a victory for marketing.
And that value can never be underestimated.

However, in the Tech Industry, Order Matters

The issue is ‘what should come first’.

In the tech industry, marketing must sit atop technology.
The reverse ultimately becomes deception.

If you raise market expectations with spectacular showmanship,
but the actual product can’t meet those expectations?

It attracts attention at first,
but eventually loses trust and disappears.

In contrast, when marketing sits atop deep technology,
only then is ‘true innovation’ delivered to the world.

Conclusion: Ultimately, Depth Wins

Spectacular marketing gathers the masses,
but overwhelming engineering depth retains people.

Of course, OpenAI’s ‘showmanship’ wasn’t meaningless.
They proved AI’s potential to the world and explosively expanded the market pie (Time-to-Market).

But when the market enters maturity, the rules of the game change.
Now it’s a fight over who can provide services more efficiently and sustainably.

Meanwhile, Google
quietly designed chips not for sale but purely for internal use behind the stage, refined compilers, and connected datacenters.

Despite being ridiculed, they didn’t abandon their ‘full stack formula’.

And November 2025.

Beyond value (2.0) and viability (2.5),
‘Gemini 3.0 Pro’ is proving the result as a complete form.

As a heavy user spending hundreds of dollars monthly on AI services, what I see is this:

Rather than being trapped in a competitive frame worrying about which aspects to gain advantage in, laying down an entirely new playing field and solidifying the foundation.

I believe this is the real game in the tech industry.

What’s easily obtained cannot be differentiation.

Not at the level of replicable ‘ideas’,
but results condensed with physical time and actual capability.

Competitiveness isn’t something you secure.
It’s secured from within.