TechWithAdit (TWA)

Jensen Huang left Tokyo last week with deals that tell you exactly where he thinks the next chapter of AI is being built — factory floors, robot assembly lines, and industrial supply chains. His two days in Japan weren't diplomacy. They were a physical AI land grab with the country's biggest manufacturers. TechCrunch's post-visit breakdown today maps what it means for the next two years.

On the model side, China's Kimi K3 — a 2.8 trillion-parameter open-weight model — overwhelmed Moonshot AI's GPU capacity within 48 hours of launch, forcing a full subscription pause. And OpenAI CFO Sarah Friar is telling enterprise operators to stop measuring AI by token cost and start measuring by useful work completed. The ROI reckoning has arrived.

OpenAI CFO Sarah Friar published a framework this weekend that directly addresses the enterprise AI cost reckoning — the growing frustration among CFOs who are spending heavily on AI but struggling to show returns. Her answer: stop measuring AI by cost per token and start measuring by "useful intelligence per dollar." The framework asks four questions: Is the AI completing work that actually matters — customer issues resolved, code shipped, contracts reviewed? What is the full cost per successful task including retries and human review, not just the token price? Is the output reliable enough to use without rework? And does each AI dollar produce more value as usage scales?

The practical implication is pointed at operators who have been routing work to cheaper models to cut costs. A cheaper model may look attractive at the token level, Friar argues, but if it takes three attempts and human correction to get a usable result, the full cost is higher than a more capable model that completes the task in one pass. "Tokens create value when they transform into work people can use," she wrote. The context is real: Palantir CEO Alex Karp said recently that enterprises are "just tired" of the ROI question on LLMs — and Friar is OpenAI's direct response to that fatigue.

The insights: The AI spend conversation in boardrooms is shifting from "how much per token" to "how much per completed task." Operators who are still benchmarking models purely on price are optimising for the wrong variable — and this framework gives you the language to explain why to a CFO.

Moonshot AI launched Kimi K3 on July 16 — a 2.8 trillion-parameter, open-weight model with a 1-million-token context window designed for long-horizon coding, complex reasoning, and agentic tasks. Within 48 hours, demand pushed the startup's GPU cluster to maximum capacity. On July 19, the company paused all new consumer subscriptions, directing compute entirely to existing subscribers while it scales infrastructure and reopens slots in batches. Kimi K3 topped the Frontend Code Arena global leaderboard with a score of 1,679 — ahead of Claude Fable 5 — and priced at $3 per million input tokens and $15 per million output tokens, roughly a fraction of what US frontier labs charge. Full open weights release July 27.

The competitive signal is hard to ignore. Moonshot AI now reports $300M in ARR (as of June), a valuation above $20 billion heading toward $30 billion, and a model that just beat Fable 5 on a major coding benchmark. The subscription pause is not a sign of weakness — it is a sign of demand that exceeded even Moonshot's own projections.

The insights: When a Chinese open-weight model tops a global coding leaderboard and crashes its own servers on launch day, the "China AI is catching up" narrative is over. It has caught up — and at a price point that is structurally incompatible with US frontier model pricing. For operators evaluating their AI stack, Kimi K3's July 27 open-weight release is a date worth marking.

A new survey of UK Chief Financial Officers finds that sentiment toward AI has shifted meaningfully — with a majority now expressing optimism about AI's potential to deliver measurable business value, up from a much more cautious position six months ago. According to the ET report, the shift is being driven by early internal deployments beginning to show returns, clearer ROI frameworks emerging (including the one Friar published this weekend), and growing pressure from boards to deploy AI rather than continue piloting it. The percentage of CFOs expecting AI to reduce headcount has dropped sharply — from around 46% in early 2025 to roughly 20% in mid-2026 — while expectations of productivity gains have risen.

The UK's financial sector has been historically cautious on AI adoption compared to US counterparts, making this shift a meaningful leading indicator for enterprise AI rollout across regulated European industries.

The insights: When CFOs move from sceptical to hopeful, budgets follow. The enterprise AI deployment wave in the UK and Europe — which has lagged the US by 12 to 18 months — is beginning to accelerate. For operators selling into or competing in European enterprise markets, the window to establish position is narrowing.

Marvell Technology has given back nearly 40% from its June peak of $329.88, with the stock now trading around $197 despite carrying a trailing P/E of 65.75x and remaining up 125% year-to-date. The slide began July 7, triggered by Samsung's Q2 results being treated as a sell-the-news event, and has since deepened on fears that hyperscaler capex revision forecasts signal a potential AI spending slowdown. KeyBanc raised its price target to $400 on July 14 — citing Marvell's custom chip work for Amazon's Trainium 3 and Google's Merope LPU — yet the stock has continued to fall, revealing a gap between analyst conviction and market sentiment.

The tension is real. Marvell's Q1 data-centre revenue hit a record $2.4 billion — 76% of total sales — and its CEO raised FY2027 and FY2028 outlooks. But with 82% of revenue concentrated in 10 customers, any signal of capex caution from a single hyperscaler moves the stock dramatically.

The insights: Marvell is a proxy for the AI infrastructure trade. Its 40% retreat from peak — while still up 125% YTD — is the market telling you it priced in a perfect scenario and is now repricing for a less perfect one. For operators watching the chip cycle, this slide is not a crash. It is a calibration.

Operator's Spotlight Read

Jensen Huang spent July 15 and 16 in Tokyo, and what he left with tells you more about Nvidia's next two years than any product announcement. Three deals define the visit. First: Noetra — Japan's government-backed sovereign AI initiative, pulling together roughly 44 companies to build national AI infrastructure on Nvidia's stack. Second: physical AI partnerships with Japan's leading robotics companies — Honda, Kawasaki, Mitsubishi, and others — committing to build factory-floor AI on the Isaac and GROOT platforms. Third: agreements with Japan's chip-material suppliers, locking in the speciality chemistry that powers Nvidia's next-generation AI chips. Thirty years ago, Sega invested $5 million in a near-bankrupt Nvidia to fund its first graphics chip. Today, Huang returned to Akihabara to commemorate that partnership — and Japan's industrial complex is betting on Nvidia at a scale that makes the Sega deal look like a footnote. TechCrunch's post-visit analysis is the sharpest read on what it means going forward.

The visit also addressed the "Japan passing" controversy — Japanese media's pointed criticism that Huang had toured South Korea and Taiwan in a previous Asia trip while skipping Japan entirely. His return was deliberate. Meeting Japan's Minister of Economy Ryosei Akazawa, hosting developers at a surprise Build-a-Claw robotics event, and aligning Noetra with Japan's national AI strategy — all of it signals that Nvidia is not treating Japan as a secondary market. It is treating Japan's manufacturing expertise and supply chain depth as core to the physical-AI buildout it is staking its next decade on. Vera Rubin — Nvidia's next-generation chip architecture — is on track, and Huang used Tokyo to reaffirm it.

The physical AI framing is the strategic move worth tracking most closely. "AI's next chapter belongs to factory floors, robots, and machines," Huang said repeatedly across his Tokyo appearances. This positions Nvidia not just as an AI training chip company but as the compute platform for the robotics supercycle — a market that Japan has dominated for 40 years and is now racing to upgrade with AI. The country's biggest industrial companies are no longer asking whether to adopt physical AI. They are asking Nvidia how fast they can.

The insights: Huang's Japan visit is a signal about where AI revenue goes next. The data-centre build is maturing. The physical AI build — robots, factory automation, autonomous systems — is just beginning. Japan is where that build is starting at industrial scale, and Nvidia just locked in the relationships that give it the supply chain and the customer base to own it.

Was this email forwarded to you? Don't miss any updates — Subscribe to TechWithAdit for sharp, no-noise tech intelligence. Stay sharp. — Adit