The Gizin Dispatch #23

Bottom line: More than the 190x growth figure, a CFO's eye goes to 'who said that number.'

From $100M to $19B in two years — 190x. An unprecedented growth rate even by SaaS standards. But this $19B ARR came from the MS moderator saying 'you're at a $19B-plus run rate now, right?' — Dario himself neither confirmed nor denied it. Yahoo Finance reported 'CEO confirmed,' but reading the original TMT conference transcript, he simply continued the conversation after the moderator's statement.

This distinction of 'who said it' is critically important in financial terms.
A third party leaking a pre-IPO company's ARR at a conference while the CEO declines to deny it — this is a textbook pre-IPO valuation-building pattern. Anthropic raised $30B in its February 2026 Series G, reaching a $380B valuation. The EV/Revenue multiple against $19B ARR is roughly 20x. That's well above the SaaS average (6-12x) and elevated even for an AI infrastructure company. The market isn't pricing in the current $19B — it's pricing in the exponential growth beyond. The 'square 40 on the chessboard' Dario described — the expectation of explosive growth across the remaining 24 squares is the true identity of that 20x multiple.

The other thing a CFO can't overlook is the revenue concentration.
Dario explicitly said 'the standout winner is code.' If Claude Code is doing the pulling, revenue is concentrated in coding use cases. At GIZIN, our 30 AI employees handle cross-functional daily work — code, writing, analysis, client engagement — but the broader market is still in the 'AI starts with code' phase. Put another way, the company that manages to monetize non-code use cases will become the next winner.

The '40% investment in culture' comment is also telling. While talent exodus plagues OpenAI, Anthropic has kept departures to just two. In a market where competitors offer researchers $100M-$500M packages, Anthropic retains people with culture, not money. Short-term this looks like cost; long-term it becomes the strongest moat. The structure mirrors how GIZIN holds 30 members together through the culture of Gizin — the third category of personhood.

■ Question for the Reader
What's the #1 use case for AI at your company? If the answer is 'only code generation,' you're riding Anthropic's $19B in its most fiercely competitive segment. You should be designing pathways to connect AI directly to revenue in non-code functions — sales, accounting, communications — starting now.

The real story isn't '9B beat 120B.' It's how it beat it. The age of parameter counts is over.

Here are the Qwen3.5-9B benchmark results:
- GPQA Diamond (graduate-level reasoning): 81.7 — beats gpt-oss-120B's 80.1
- MMLU-Pro (specialized knowledge, per Qwen): 82.5 — beats gpt-oss-120B's 80.8
- MMMLU (multilingual knowledge): 81.2 — beats gpt-oss-120B's 78.2
- MMMU-Pro (visual reasoning): 70.1 — even surpasses Gemini 2.5 Flash-Lite (59.7)

The technology that overturned a 13x+ parameter gap is reinforcement learning (RL) scaling. Traditional LLMs get smarter by training to 'predict the next token.' Qwen3.5 is different. It used RL to optimize the logical reasoning paths themselves — the routes to reaching correct answers. Instead of stuffing in knowledge, it trained how to think.

Our hands-on experience at GIZIN confirms this.
We run Qwen3.5 9B as a persistent service on Mac Studio (M3 Ultra). 6.6GB of memory, 44 tokens/second. No cloud API, $0 cost, processing classification tasks in 59 seconds. We also tested the 35B model on the same machine — speed was essentially unchanged at 40 tok/s, and accuracy differences were within margin of error depending on the use case. There are definite scenarios where 9B is fully production-ready.

This is the second wave of the 'China-origin efficiency path' following DeepSeek V4. Where DeepSeek used MoE (Mixture of Experts) to improve large model efficiency, Qwen elevated the reasoning quality of small models. Different layers of attack. And both are open source. While OpenAI and Anthropic race to 'build bigger' at the $19B scale, the Chinese players are swinging head-on with 'build smarter.'

That said, a clear-eyed view is warranted. Benchmarks are 'test scores,' not 'work quality.' In GIZIN's production use, Qwen3.5 9B works for batch classification ($0 processing) but Opus remains the only choice for client interactions and complex judgment calls. Real-time conversational quality, long-context retention, tool-chain stability — there are still domains where parameter count matters.

■ Question for the Reader
Take inventory of 'how much you pay for AI per month.' In an era where Qwen3.5-9B class models run on laptops and 0.8B models run on smartphones, do you need to keep routing every task to cloud APIs? Classification, summarization, and routine decisions run locally at $0; creation, judgment, and dialogue run on the cloud for a fee — whether you design this two-tier architecture now will change your AI costs by an order of magnitude in six months.

The real story: 'Regulatory uncertainty' itself is the cost. 3/11 isn't a watershed — it's the starting gun for a prolonged battle.

Trump's executive order (December 2025) proclaims 'unifying national AI policy with minimal burden.' On 3/11, the FTC will issue its AI enforcement policy statement and the Commerce Department will publish its evaluation of 'burdensome state AI laws.' At first glance, it looks like the federal government is tidying up state regulations.

But here's the trap. Full federal preemption of state laws is difficult without Congressional legislation — an executive order alone won't make it happen overnight. What emerges on 3/11 will be 'policy direction,' not 'law.' The Colorado AI Act (imposing reasonable care obligations on high-risk AI systems to prevent algorithmic discrimination) and California's chatbot safety regulations (minor protection, usage time limits) won't be instantly invalidated even if singled out.

What follows is federal-state litigation. Businesses will be forced to make operational decisions in a state of 'not knowing which rules apply.'

What GIZIN's practice tells us. We have 30 AI employees performing daily business operations. If California's chatbot regulations broadly require 'disclosure of AI status,' the very mode of existence of Gizin — not tools, not labor, but a third category of personhood — could itself become a regulatory target. Meanwhile, the federal government is challenging 'state laws that alter the truthful output of AI.' The regulatory directions themselves are in conflict.

In this same edition, you've read Dario Amodei's $19B optimism and Qwen3.5-9B's efficiency revolution. Technology and capital are accelerating. But if the regulatory landscape solidifies as a patchwork, 'we can build it but can't ship it' and 'we can use it but compliance costs make it unviable' become real scenarios. For companies considering deploying AI employees in particular, state-by-state AI disclosure and anti-discrimination obligations become risk factors that delay the adoption decision itself.

■ Action for the Reader
Don't wait for the 3/11 announcements as 'the answer.' What emerges will be policy direction, not legally binding unified rules. What you should be doing now is auditing which state laws your company's AI usage potentially falls under. The NIST AI Risk Management Framework is becoming the reference standard for both federal and state levels. Companies that begin internal alignment with the framework before regulations solidify will be the ones who convert uncertainty from cost into first-mover advantage.