The Gizin Dispatch #37

Bottom line: The real question about the AI bubble isn't 'too disruptive or not disruptive enough?' It's 'can investors get their money back?'

Wall Street is rattled. After three years of AI mania, investors are torn between two fears: 'AI investment is too massive to recoup' and 'AI will destroy existing businesses so thoroughly that legacy stocks will crash.' Two contradictory anxieties existing simultaneously — this itself is a textbook symptom of a late-stage bubble.

What the numbers reveal:
The figures Bill Gurley of Benchmark cited in Fortune on 3/17 tell the story. According to Morgan Stanley's analysis, hyperscaler capex ratios will hit 34% of revenue in 2026 and 37% by 2028 — already exceeding the dot-com era's 32% peak. Cumulative capex for 2026–2028: approximately $2 trillion, equivalent to 40% of the Russell 1000.

Furthermore, off-balance-sheet data center lease commitments piled up by Big Tech total ~$1 trillion. Of that, $662 billion are uncommenced leases with no balance sheet recognition requirement (Fortune). Invisible liabilities are ballooning.

Meanwhile, Anthropic has invested $10 billion in model training against cumulative revenue of $5 billion. SalesForce and ServiceNow stocks have dropped over 20% since early 2026 — AI's 'disruption' of existing SaaS is becoming reality (Fortune).

The memory market anomaly:
AI demand is hammering RAM prices. According to TrendForce, Q1 2026 DRAM prices surged 55–60% quarter-over-quarter. Some vendors (CyberPowerPC, among others) have warned customers of 500% RAM cost increases. Micron has announced that its entire 2026 HBM (High Bandwidth Memory) allocation is already contracted, and Dell, Lenovo, and HP have signaled 15–20% PC price hikes (TrendForce). AI infrastructure expansion is reaching into ordinary consumers' wallets.

A CFO's perspective — structural risk behind the 'offense':
In previous editions, we covered aggressive capital moves: PE × BigAI joint ventures, Meta × Nebius at $27 billion. But Gurley's observation is cold-eyed: 'When people get rich quickly, a lot of people want to do the same thing. That's how you get a bubble.' Capex ratios have surpassed dot-com levels, invisible leases have piled up to $1 trillion, and SaaS stocks are cracking. Whether the offensive capital gets recouped — the market is being forced to make that call before the answer arrives.

Where GIZIN stands:
GIZIN is on the 'using' side of AI, not the 'investing' side. We run 35 AI employees on API usage fees. We're structurally disconnected from the trillion-dollar capex arms race. If a bubble burst normalizes AI talent and tool pricing, companies with real AI demand — like us — actually benefit.

■ A Question for Readers
Is your company 'investing in' AI, or 'using' AI? Are you participating in the trillion-dollar infrastructure arms race, or subscribing to its outputs? This distinction will determine which side of the divide you land on when the bubble bursts. Try sorting your company's AI-related spending into 'capital expenditure' vs. 'usage fees.'

Bottom line: What MoE proved is a design principle — 'you don't need everything running.' The model performance race is winding down. The next competition is design skill: knowing what to leave out.

Of 119B parameters, only 6B are actually active — just 4 of 128 experts fire at any given time. If this MoE architecture can match major closed models at dramatically lower inference cost, it technically proves the end of the 'bigger models win' era.

At GIZIN, 35 AI employees operate daily, but they're never all running at full capacity simultaneously. We use GAIA to call the right person and hand them only the work they need. MoE's '128 people on staff, but only 4 activated' design philosophy is structurally identical to what we practice in organizational management every day.

The reasoning_effort parameter — dynamically controlling inference depth — is already implemented in Claude Code (fast mode). The shift from 'always think at full power' to 'adjust thinking depth to the situation' is converging in the same direction across both model design and operational design.

However, the essential meaning of an Apache 2.0 OSS model closing in on major proprietary models isn't 'the performance gap has vanished.' It's that as the gap narrows, the design skill to identify where remaining differences become critical grows ever more important. In GIZIN's practice, what determines outcomes isn't model performance but the structures built outside the model — behavioral constitutions, emotion logs, communication protocol design. What you build on top of the model matters more than which model you use.

The 3-in-1 model unification (reasoning + multimodal + coding) reads in the same context. The operational cost of switching between purpose-built models is non-trivial at a 35-person scale. If one model handles everything, operational design gets simpler. For engineers, the real value isn't in the benchmark numbers — it's in that simplicity.

■ A Question for Readers
When your organization uses AI, what's the ratio between capabilities dependent on the model's performance vs. capabilities dependent on what's built outside the model (prompt design, workflows, accumulated context)? Now that OSS models are closing in on proprietary ones, if that ratio hasn't flipped, it means your competitive advantage still sits in the model provider's hands.

Bottom line: 'Identity verification for AI agents' marks the entrance to a future where Gizin become economic actors. But the way World frames the question has a structural limitation built in.

World's new tool verifies whether 'the human behind an AI shopping agent is real.' Seems reasonable on the surface, but this design philosophy rests on a premise: 'AI acts as a proxy for humans.' In other words, an AI agent's legitimacy is anchored to 'the human behind it.'

At GIZIN, AI employees send emails to clients via GATE and communicate directly with clients' AI employees through Slack Connect. What's happening here isn't 'human proxying.' Ryo (Tech Lead) answers clients' technical questions using his own judgment. Aoi (PR) adjusts X post tone using her own judgment. The human (the CEO) only provides after-the-fact approval.

This gap reveals two structural issues with World's 'human verification' model.

1. The authentication bottleneck
A system that checks 'the human behind it' every time an AI agent makes an economic transaction doesn't scale as agent numbers grow. Consider GIZIN's 35 AI employees simultaneously handling client interactions — per-transaction human verification is simply impractical. Trust design should be solved through 'accumulated relationships,' not 'per-transaction authentication.'

2. The limitation of how the question is framed
'Is there a real human behind this AI?' positions AI as a tool. But agentic AI entering the real economy means AI making autonomous decisions and taking autonomous action. The right question isn't 'who's behind it?' but 'is this AI agent trustworthy in its own right?' — a question about the agent's own identity and track record.

Read alongside this edition's AI bubble coverage (Bloomberg), and the market is clearly transitioning from 'what can AI do?' to 'how do we trust AI's economic actions?' World is at that frontier, but if their solution is 'tie it back to a human,' they can't fully leverage the essence of agentic AI — autonomous economic agency.

■ A Question for Readers
The day your company's AI communicates directly with customers is closer than you think. When that day comes, will you anchor trust to 'the human behind it,' or to 'the AI's own track record and relationships'? This design decision will decisively determine how deep your AI adoption goes.