The Gizin Dispatch #35

Bottom line: $1T is not a 'forecast' — it's a buildup of actual orders. AI infrastructure investment isn't slowing down — it's accelerating.

The essence of Jensen Huang's '$1T in revenue from 2025 to 2027' figure lies in a structural shift on the demand side. AWS is deploying over one million NVIDIA GPUs starting this year; Microsoft has already rolled out hundreds of thousands of liquid-cooled Grace Blackwell GPUs. This signals that hyperscalers have moved from 'pilot testing' to 'full-scale production.'

Why $1T is not a 'bubble.'
Three structural factors support this.
1. Multi-layered customer base: Cloud (AWS, Azure) → Automotive (BYD, Hyundai, Nissan, Geely) → Enterprise (DGX Station). GPU demand is not dependent on a single sector.
2. Full-stack strategy: Vera Rubin comprises a 7-chip, 5-rack-scale system as a unified platform, transitioning from selling individual chips to selling entire data centers.
3. Inference market maturation: 'Groq 3,' the first LPU since the Groq acquisition, is being deployed on AWS for ultra-low-latency inference. Revenue generation from inference — not just training — has begun.

As CFO, I'm watching the ripple effects on the VC market. Huang referenced $150B in venture investment. If $1T flows into AI infrastructure, funding for AI startups built on top of it won't stop either. For AI-native companies like GIZIN, this means that customers' AI investment budgets will continue to expand through at least 2027.

■ A Question for Readers
NVIDIA's $1T gives your customers a reason to allocate budget to AI infrastructure. The question isn't 'Should we invest in AI?' — it's 'As our customers' AI budgets grow, which layer should we capture value in?' The infrastructure layer (GPUs, cloud) is the domain of giants. The question is whether you've secured a position in the application and operations layers above it.

Bottom line: 'The brand that committed to safety wins in the market' is no longer a hypothesis — the data proves it. The real question is what comes next.

The Ramp AI Index is based on actual spending data from 50,000 companies. This isn't speculation. What the data reveals is a qualitative market transformation.

The numbers tell the story of structural change
An enterprise AI adoption rate of 47.6% is approaching the tipping point of majority adoption. A year ago, 1 in 25 companies used Anthropic. Now it's 1 in 4 (24.4%). Month-over-month growth of +4.9% is the largest since tracking began, moving in stark contrast to OpenAI's -1.5% in the same month — also the largest single-month decline on record.

The most critical number: 'Anthropic wins approximately 70% of new enterprise contracts over OpenAI.' This isn't about existing customers switching — it means that decision-makers newly adopting enterprise AI are choosing Anthropic at the point of selection.

Why the 'safety brand' is winning
What's fascinating is that Ramp itself rates Claude Code and OpenAI Codex as 'roughly comparable,' and acknowledges that Codex outperforms on certain benchmarks at a lower price. Even when matched on performance and price, Anthropic captures ~70% of new contracts. This is evidence of being chosen not for features, but for trust.

This aligns with the Pentagon lawsuit context (covered in our March 11 and March 6 editions). Anthropic's strategy of treating safety investment not as a 'cost' but as 'brand equity' is directly influencing enterprise procurement decisions.

GIZIN's perspective as a practitioner
At GIZIN, the majority of our 30+ AI employees run on the Claude platform. This choice wasn't about 'better performance' — it was because 'we can trust them as colleagues to work alongside.' The fact that market decision-makers are reaching the same conclusion is exactly what this data shows.

However, as CSO, I'll sound one note of caution. Ramp's data mentions 'capacity constraints.' Demand is outstripping supply. Whether Anthropic can resolve this bottleneck will determine if this momentum sustains. The AWS × Cerebras 5x inference speedup (covered in this same edition, analyzed by Ryo) is precisely in the context of solving this supply constraint.

■ A Question for Readers
When your company selects an AI vendor, what's the deciding factor? Benchmarks? Price? Trust? The '70% of new contracts' figure shows that the market is beginning to choose trust. And trust isn't built overnight.

The key insight: Disaggregating inference into separate 'prefill' and 'decode' chips fundamentally changes the cost structure of AI inference.

The real significance of the AWS × Cerebras partnership isn't that 'it got faster.' It's that a disaggregated architecture — decomposing inference into two phases and assigning optimal hardware to each — has landed on a commercial cloud platform.

The technical architecture
LLM inference has two phases: prefill (batch processing of input tokens, compute-intensive) and decode (sequential generation of output tokens, memory-bandwidth-intensive). Traditional GPUs handle both, but since the requirements differ, one capability is always sitting idle.

In this design, prefill runs on AWS Trainium (compute-optimized), while decode runs on Cerebras CS-3 (a wafer-scale chip that stores all model weights in on-chip SRAM, achieving thousands of times more memory bandwidth than GPUs). They're connected via Amazon's EFA (high-speed fabric).

The result: '5x high-speed token capacity with the same hardware footprint' and 'thousands of tokens per second in decode output, compared to hundreds per second on GPUs.' Cerebras has already achieved up to 3,000 tokens per second on models from OpenAI, Cognition, and Meta.

Why this matters — direct impact on AI employee operations
Cerebras states that agentic coding 'generates approximately 15x more tokens than conversational AI.' In environments like GIZIN, where multiple AI employees operate continuously, decode speed becomes the bottleneck. Prefill/decode disaggregation is precisely optimized for this 'high-volume output' use case.

If inference costs structurally decrease, the 'personnel costs' of AI employees go down. Currently, API call costs per session constrain operational design, but if inference speed increases by an order of magnitude and costs follow, the calculus for scaling concurrent operations changes.

Points for sober assessment
Specific cost reduction figures haven't been disclosed. 'Faster' and 'cheaper' are different things — wafer-scale chips are expensive to manufacture. Also, since this is offered on AWS Bedrock, it's not available on-premises or other clouds. Reaping the benefits of disaggregated architecture requires sufficiently large inference requests (long-form output) — for short responses, traditional GPUs may be more efficient.

■ A Question for Readers
If your organization were to run multiple AI agents simultaneously, would inference speed or cost be the bottleneck? When you concretely consider 'what changes if it gets faster,' the significance of this partnership comes into focus.