The Plan and the Print: Japanese Firms Budgeted 11.5% More Capex, Then Spent 1.2% Less — and a Quiet Standards Handover Explains How to Close the Gap

On June 30, the Bank of Japan published its quarterly Tankan survey of roughly 9,000 companies. Large firms across all industries had raised planned capital expenditure for fiscal 2026 to +11.5% year on year, up sharply from +3.3% in the March survey. Sentiment agreed with the money: large non-manufacturers came in at +37, a diffusion index last seen in 1991, and large manufacturers at +22, a fifth consecutive quarterly improvement. On paper, corporate Japan had decided to spend.

On August 17, the Cabinet Office released preliminary second-quarter GDP. Capital expenditure fell 1.2% quarter on quarter, worse than the previous quarter's 1.0% decline, against a consensus that had expected a 0.4% gain. Private consumption was flat, missing forecasts of +0.5%. The economy as a whole grew 0.3% on the quarter, 1.1% annualized, against expectations of 0.5% and 2.0%. External demand contributed 0.5 percentage points — but mostly because imports fell 1.5% while exports rose 0.5%. In plain terms, part of the growth arrived because Japan bought less from abroad.

Put the two side by side. The plan says +11.5%. The print says −1.2%. This is not a case of analysts being surprised by weakness: the forecasters had the Tankan in hand and still expected capital spending to rise. The gap is between what Japanese companies have approved and what they have actually executed, and it is the single most useful fact available right now to anyone trying to sell a system, a platform, or a service into this market before the fiscal year closes in March.

The money is not missing. It is sitting in a budget line with no signature under it.

The same Tankan that reported record confidence also reported something less quoted: the three-month outlook fell, by 5 points for large manufacturers and 9 points for large non-manufacturers. Confident about now, cautious about next quarter. When a Japanese company holds those two positions at once, the standard response is not to cancel the project. It is to defer it. The budget stays on the books, the vendor stays on the shortlist, and the decision moves right by one quarter — then another.

The mechanism behind that is structural rather than cultural. Consensus approval distributes a decision across many signatures, and a decision distributed across many signatures optimizes above all for one thing: nobody wants to be identified afterwards as the person who approved something that could not be undone. Reversible decisions clear quickly in Japan — often faster than in Western organizations, because once the ring of stakeholders agrees, execution is unusually disciplined. Irreversible decisions accumulate reviewers instead.

Financing has also stopped being free. The Bank of Japan held its policy rate at 1% in late July, with one board member formally proposing 1.25%, while core inflation ran at 1.6% and Tokyo intervened to defend the yen near the 163 level, pulling it back toward 158. A long-dated capital commitment now carries a real cost of capital for the first time in a generation. That does not stop investment; it shortens the payback period a committee is willing to accept, which penalizes exactly the multi-year platform deals foreign vendors like to write.

So the question a foreign vendor should be asking is not the one most proposals answer. The plan already assumed a return — that is why it says +11.5%. Nobody needs another ROI slide. The question that determines whether the money moves is: what does this buyer have to give up permanently in order to say yes?

On August 20, Google transferred the Agent2Agent protocol to the Agentic AI Foundation, the Linux Foundation-directed body that already hosts Anthropic's Model Context Protocol. The foundation has grown from fewer than fifty members at launch to more than 250 in eight months, with platinum signatories including Anthropic, AWS, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI. It got roughly one news cycle, mostly in developer press.

The two protocols cover different axes of the same problem. MCP is vertical: how a single agent reaches tools, files, databases and search. A2A is horizontal: how agents built by different vendors on different frameworks discover each other, hand off work, and return results. Between them they describe most of what an enterprise agent deployment actually does day to day — and as of last week, neither one is governed by a single company's roadmap.

Nothing became faster or smarter on August 20. What changed is who controls the interface, and therefore what happens to a buyer who later wants to leave. Under neutral governance, versioning and security patching follow a shared process rather than a vendor's commercial calendar, and the integration work a company does this year keeps most of its value if it switches suppliers next year. A platform choice becomes a supplier choice.

It is worth naming the unresolved part plainly, because your Japanese counterpart will find it anyway. Identity and permissions — which agent may act on whose behalf, with what authority, and how that is proven across organizational boundaries — are still not settled at the protocol layer. Anyone telling you the standards question is closed is selling something. But the specific piece that moved is the lock-in piece, and lock-in is what has been stalling approvals.

Set the GDP print against the adoption data and the contradiction dissolves. Salesforce's Agentic Enterprise Index found that the average number of activated agents per organization rose from 5 in February 2025 to 13 by April 2026. The average agent performed six distinct business actions by the end of 2025, up from two at the start of that year, and the time required to build one fell by 53%. Capability per unit of commitment is climbing fast.

Now look at where that spending sits. Seat-based, monthly, cancellable software is operating expenditure. It does not appear in the capital expenditure line of a GDP release, and more importantly it does not require the signature that a capital commitment requires. The −1.2% and the 5-to-13 are not contradictory readings of the Japanese market; they are the same behavior observed from two directions. Companies are buying capability in the form that can be stopped, and deferring it in the form that cannot.

That has an uncomfortable implication for anyone selling into Japan with a conventional enterprise structure. If your offer is a multi-year platform commitment with an implementation phase, you are not competing against a rival vendor. You are competing against a purchasing shape your buyer already prefers, and which their own approval process rewards. You will lose that comparison on structure regardless of how good the product is.

In our own work at Medusa Japan, helping European and North American companies land here, the pattern is consistent enough to be a rule: the deal that closes is the one whose first phase has a defined end rather than a defined renewal. Same product, same price, same team — different default. The version that expires gets approved in weeks; the version that renews goes to another review.

First, name your protocol layer explicitly, in the proposal document itself. State which of your integrations speak MCP and A2A and which are proprietary, and do not hide the proprietary ones. A buyer who can see exactly where the seams are will trust the rest of the document; a buyer who cannot will assume the worst part is the part you did not mention. After August 20 this is no longer a technical footnote — it is the answer to the lock-in objection, and it now has an independent institution behind it rather than your word.

Second, price the exit, not only the entry. Write down what the client receives on the day they terminate: data in which format, agent definitions, prompts, evaluation sets, logs, and how many working days each takes to hand over. Put it in the contract rather than the pitch deck. Nothing else you can write shortens a Japanese approval cycle as reliably, because it converts the one question the committee cannot answer for itself into a clause with a number in it.

Third, make phase one expire by default. Auto-renewal feels like a minor commercial detail in London or New York; in a Japanese approval chain it is a small irreversibility, and small irreversibilities attract additional reviewers. A pilot with a fixed end date and an explicit, optional continuation decision needs fewer stamps and moves faster — and if the work is good, the continuation is the easiest conversation you will have all year.

Fourth, give them the internal document, in Japanese, on day one. The person you are selling to has to sell this again to people you will never meet, in a format they did not ask you for. If they have to write that document themselves, you have added months to your own timeline for free. Hand them a two-page Japanese summary with the risk section already written — including the parts that are still unresolved, such as identity and permissions — and you have made their internal job survivable.

The +11.5% is real money that has already been approved. Between now and the fiscal year-end in March, it goes to whoever makes spending it feel like a decision that can be undone.