Broadcom Delivers the Check, MUFG Delivers the Client: Who Funds the Next Wave of AI in Japan

Broadcom reported on September 2nd a revenue of $16.7 billion in AI-focused chips for the quarter ending in August, a 221% increase year-over-year. The figure has been interpreted in the market as confirmation that the demand for AI infrastructure remains tied to a few hyperscalers, but there is a second, more discreet layer that will define the next cycle: who pays for the consumption of these chips at the enterprise level. In Japan, this layer has three names: MUFG, Mizuho, and SMBC.

The Contract That Leaves the Datacenter and Enters the Bank

The three megabanks acquired, according to a report by Nikkei published in August, access rights to the new generation of OpenAI models. This move is complemented by MUFG's investment in Sakana AI, a Japanese unicorn in which the group has been the largest local corporate investor since 2024, maintaining a three-year development agreement signed in May 2025. In the roadmap released, MUFG is using this asset to launch a digital bank with Sakana AI and LayerX in the current fiscal year, as well as embedding credit assistants trained based on decisions from human managers.

The economic point is straightforward. According to CNBC, Anthropic reported this week an annualized revenue of over $30 billion with more than a thousand clients spending above $1 million annually, a number that has doubled in less than two months. OpenAI reports a similar trajectory. Contracts like those of the three Japanese banks are foundational to this doubling. The silicon from Broadcom is the same dollar that MUFG, Mizuho, and SMBC extract in productivity, and it is the dollar that underpins the nearly $1 trillion private valuation that investors currently attribute to each of the two major labs.

Where the Human Cost Appears Differently

The standard reading from the West translates the cycle into workforce reduction: Standard Chartered announced 7,800 fewer positions by 2030, Morgan Stanley cut about 2,500 jobs, and Bank of America says its workforce will shrink in 2026 due to unfilled attrition. Applying the same model to Japan miscalculates on two fronts. Lifetime employment and internal relocation limit layoffs as a tool. Moreover, the Japanese banking sector is already facing structural talent shortages in areas like SME credit and compliance due to demographic aging. AI does not replace jobs that are closing; it fills positions that banks cannot find candidates for.

The counterpoint comes from the sector itself. Masahiro Kihara, CEO of Mizuho Financial Group, presented in February a plan to replace around 5,000 administrative positions with AI over ten years, stating, "I do not believe that humans will lose value," as they "can seek higher-value-added work." From MUFG's side, Hironori Kamezawa framed the goal in different terms: "We need to transform into an AI-native company." Both statements describe a project of reconfiguration, not of aggressive cuts. If they are correct, the effect on the workforce will manifest over the years, in the makeup of the staff, not in a wave of layoffs.

What Weakens the Thesis

Two pieces of data weaken the optimistic reading. First, Broadcom guided a consolidated gross margin of 73% for the fourth quarter, down from 78% a year earlier. The 5-point drop indicates that the price of AI capacity is under pressure, and part of what is currently gross revenue may become a commercial concession to the next enterprise client. Second, the pace of launches from the labs, which CNBC classified as model fatigue on September 6, shifts hidden costs to the Japanese buyer. Each new version forces MUFG to revalidate the digital bank pilot, and local regulatory approval does not keep pace with the weekly cycle.

The distinction that rarely appears in the debate is between a hyperscaler burning venture capital and a bank with over $3 trillion in assets funding AI with its own operating profit. MUFG's account does not depend on the next funding round to pay for Broadcom's chips. It derives from the banking spread, foreign exchange, and sovereign portfolio. It is a thesis different from the one haunting the West, where the same capex is justified by the future promise of a reduction in white-collar jobs.

Two concrete milestones dictate the next cycle. The launch of the MUFG-Sakana-LayerX digital bank, expected in the current fiscal year, will test in practice whether a system designed with AI at its core can achieve a lower acquisition cost than the second-generation digital banks in the country. And the next earnings report from Broadcom, due in early December, will provide material to separate the dollar that came from hyperscalers from the dollar that came from buying banks. If margin dilution persists and the mix remains concentrated, this quarter's 221% growth will be viewed as a peak, not a floor.