EVERY PROMISE
EVENTUALLY MEETS
A BALANCE SHEET.

Stonkholders,

There was a time when management could identify the nature of an asset by looking at what happened when markets became uncomfortable.

Gold offered protection. Technology stocks offered growth. Bitcoin mostly offered more Bitcoin. Treasuries rallied when the economy weakened.

Those categories have become less cooperative.

This week, crypto assets increasingly moved alongside gold and silver. At the same time, the long end of the U.S. Treasury market continued to demand yields beginning with a five, even as economic signals moved in different directions and the Treasury Department became more visible in its efforts to improve conditions in its own market.

Management does not think those observations are unrelated.

The market appears to be paying more attention to assets that do not rely on somebody else's future ability or willingness to pay. Gold, silver and Bitcoin share very little mechanically, but all three sit outside the conventional chain of financial liabilities. That distinction appears to matter more than it did a year ago.⁴⁵

BITCOIN HAS BEEN INVITED TO A DIFFERENT MEETING

Management is not prepared to declare Bitcoin a reserve asset. It is, however, difficult to ignore that the market has begun inviting it to some of the same meetings as gold.

For much of its history, crypto behaved like a highly volatile expression of global liquidity. It participated aggressively when speculative conditions improved and usually discovered the downside just as enthusiastically when liquidity disappeared.

That relationship has not vanished. What has changed is the frequency with which Bitcoin now appears alongside gold and silver during periods of concern about inflation, fiscal sustainability and the purchasing power of conventional financial claims.

The three assets are not interchangeable. Silver retains a large industrial component. Gold has thousands of years of monetary history. Bitcoin remains younger, more volatile and dependent on a very different market structure.

Their common feature is narrower: none is a sovereign promise to pay.

That matters most when investors are beginning to ask more questions about the promises already sitting inside their portfolios. The issue is not whether currencies or government debt are about to be abandoned. It is whether the desired allocation to alternatives rises when confidence in the long-term purchasing power of conventional financial assets becomes less automatic.

That question becomes more relevant when the benchmark sovereign borrower itself is being asked to pay more for duration.

THE BOND MARKET HAS REVIEWED THE BALANCE SHEET

In the previous letter, management argued that the long end of the Treasury curve appeared to be discounting something beyond the economic cycle.

Earlier labour-market and consumer data had suggested that growth might be losing momentum. The August flash purchasing managers’ survey complicated that interpretation: the composite index rose to 56.0 and services to 56.8, while manufacturing eased to 53.2. The survey pointed to third-quarter growth approaching 3% annualized, not to an economy already in retreat.¹⁸

Some of the compensation demanded by long-duration Treasuries may therefore reflect stronger activity. Management does not intend to assign the entire long-end move to fiscal supply when the growth data have offered a competing explanation.

This week provided a useful additional test.

The Treasury Department doubled planned buybacks of longer-dated securities as policymakers sought to improve market functioning and reduce pressure in parts of the curve. Bonds initially rallied. The move did not hold, and by the end of the week the 30-year yield remained above 5% after trading at levels not seen since 2007.¹²³

Treasury buybacks are not quantitative easing. The institution, objective and mechanics are different. That distinction matters.

What also matters is that pressure in long-duration government debt had become significant enough to generate a more visible policy response, yet the market continued to demand a substantial premium afterward.

The stronger activity figures make part of the move easier to understand. They do not, however, explain away the Treasury’s decision to expand liquidity-support operations or the premium still being attached to long duration. Growth expectations, fiscal supply, inflation uncertainty and term premium can all be present in the same yield.

There is an interesting contrast here. Investors are demanding more compensation to own the world's benchmark sovereign liability at the same time that assets without an issuing balance sheet are attracting renewed attention.

That is not evidence of monetary collapse. It is evidence that the market is spending more time thinking about the quality and durability of financial claims.

THE CONSUMER SUPPLIED THE FOOTNOTES

The previous letter asked this week's retail earnings to tell us whether July's weaker consumption data represented broad deterioration or something more specific.

The answer was more nuanced than a consumer recession.

Target reported stronger comparable sales and traffic. Home Depot continued to see reasonable demand for repair and maintenance. Lowe's reduced its outlook as larger discretionary projects remained weak, while Walmart continued to frame value, speed and convenience as central to household spending decisions.⁶⁷⁸⁹

The important signal is not that consumers have stopped spending. They have become more selective about where that spending goes.

Value remains attractive. Necessary purchases continue. Smaller projects can still clear the hurdle. Larger, deferrable and financed purchases appear more sensitive to the cost of money and the household budget.

That distinction matters because headline consumption can remain resilient even while the composition underneath it becomes less healthy. Higher energy costs absorb spending power without necessarily causing overall retail sales to collapse. Expensive financing delays purchases that can be postponed. Households substitute rather than disappear.

The PMI supplies a useful cross-check. Services accelerated and hiring improved even as manufacturing softened. The Federal Reserve has likewise described resilient consumer spending alongside increasing pressure on lower- and moderate-income households. Aggregate demand can improve without removing the unevenness underneath it.¹⁸¹¹

This is a less dramatic story than consumer collapse, but it fits the broader macro picture better.

THE DISINFLATION WAS RENTED

July's inflation reports were better than expected, and markets were right to respond to them.

The problem was not the data. It was the assumption that the underlying conditions would persist.

Energy helped. Energy has since stopped helping.

Oil moved sharply higher as the geopolitical situation around Iran and global supply routes deteriorated, with Brent ending the week above $90 per barrel. At the same time, Federal Reserve minutes showed that many policymakers remained concerned enough about persistent inflation to contemplate further tightening if progress toward target stalled.¹⁰¹¹

The August survey did show some moderation in input costs and selling prices. Both remained elevated, and renewed energy pressure could reverse the improvement. That leaves the Federal Reserve managing an expansion that may be stronger than last quarter while inflation remains above target and long-term borrowing costs stay restrictive.¹⁸¹¹

For that reason, management remains less interested in whether September brings another hike than in whether financial conditions actually become easier.

The two questions are no longer equivalent.

SOMEONE STILL HAS TO PAY FOR THE DATA CENTRE

Another balance sheet now deserves closer inspection.

One of the strongest arguments against a direct comparison between the present AI investment cycle and the late 1990s has been the way the buildout was financed.

The largest hyperscalers entered this cycle with enormous cash flows and some of the strongest balance sheets in corporate history. Much of the initial infrastructure expansion could therefore be funded internally. If too much capacity was built, the first-order consequence was a lower return on capital rather than an immediate credit event.

That distinction has always mattered.

It may now be becoming less clean.

Nvidia and major Wall Street firms have announced financing platforms targeting more than $500 billion of capital for AI infrastructure. Other transactions increasingly involve project finance, long-term leases, guarantees, vendor support and private credit.¹²

The Federal Reserve’s July minutes separately noted that more AI capital spending was being financed with borrowing, including credit from nonbank investors and regional banks.¹¹

Reuters also relayed a Bloomberg report that Broadcom was exploring more than $60 billion of AI-related debt financing through a special-purpose structure. The transaction was reported as under discussion, not completed. Its relevance is that more of the infrastructure trade is beginning to involve creditors who do not share an equity investor’s ability to wait.¹⁹

The Wall Street Journal's analysis this week made the scale of the commitments easier to see. Across nine major technology companies, it identified more than $3 trillion of additional commitments beyond the capital expenditure already appearing in conventional reported figures, including future leases, equipment purchases and energy arrangements.¹³

That spending can lift current GDP and support employment while the assets are being constructed. Counting a data centre as investment does not establish that its eventual revenue will justify the capacity, service the financing or support the valuations built around it.

The useful question is therefore shifting.

It is no longer just how much AI infrastructure companies intend to build. It is how those commitments eventually migrate onto balance sheets, who funds them, and who is exposed if the expected utilization does not arrive.

None of this makes the underlying demand fictitious. Debt-financed infrastructure can be productive infrastructure, and a long-term lease can support a perfectly sound business.

The concern is about how the downside behaves.

A hyperscaler that overbuilds using retained earnings can absorb a disappointing return for a long time. Once more creditors, lessors, guarantors and external capital providers enter the structure, the consequences of an error spread beyond the equity holder.

That is where the comparison with the dot-com period becomes more relevant than it was eighteen months ago.

The technologies are different. The companies are different. The lesson that survives both cycles is simpler: capital structure determines how an investment mistake travels through the financial system.

If AI demand disappoints while most infrastructure remains internally funded, investors can mark down the owners of the capacity. If leverage becomes a much larger part of the buildout, the same mistake has additional transmission channels.

Management is not filing a claim that this transition has already gone far enough to create a systemic credit problem.

It is filing a note to watch the direction of travel.

THE PROVIDER IS ALSO PERMITTED TO HAVE A BUSINESS MODEL

There is another assumption worth reviewing.

The AI cycle does not only require enough capital to build the infrastructure. It also requires enough demand to consume what gets built.

For several years, the economics of that demand have been helped by rapid declines in inference costs. Competition between providers has been aggressive, infrastructure has been heavily subsidized, and new entrants have had every incentive to prioritize adoption.

That has made it easy for enterprises to assume that whatever AI costs today will probably cost less tomorrow.

DeepSeek supplied a useful counterexample this week.

The company introduced new peak and off-peak API pricing for its V4 models. Depending on the model and token category, prices increased by between roughly 50% and 1,100%. For V4-Pro, cached-input pricing at peak increased from $0.003625 to $0.044 per million tokens, while peak output pricing rose from $0.87 to $3.96.¹⁴¹⁵

DeepSeek remains extremely competitive on price, which is exactly why the change is interesting.

The cost pressure is not limited to model pricing. Bloomberg reported that some customers had been told servers containing Nvidia’s AI chips could rise by more than 15% as memory costs increase; Reuters relayed the report but could not independently verify it. If borne out, the example would point to pressure on infrastructure costs before those systems are even deployed.²⁰

The pricing announcement does not prove that inference costs are about to rise across the industry. It does show that falling prices are not a permanent contractual feature of the market.

At some point, providers may choose to optimize for economics rather than adoption.

The enterprise implication is not obvious from the price of a single prompt.

That framing increasingly understates how AI is being used.

An enterprise agent may make repeated model calls, retrieve documents, invoke tools, review its own work, retry failed steps and ask another model to validate the result before one business process is complete. A modest change in model pricing can therefore propagate through a workflow many times before it reaches the final unit of output.

This is why the end of the subsidy era, if it is beginning, matters beyond model-provider margins.

A large amount of expected enterprise AI adoption rests on very large future usage volumes. Those volumes are easier to justify when the unit economics are assumed to improve continuously.

If providers eventually discover durable pricing power, some enterprise use cases will survive easily. Others will have to demonstrate a return on investment under more conventional economics.

That is probably healthy for the technology.

It may be less comfortable for the revenue forecasts already assuming the volume.

EVERY PROMISE EVENTUALLY MEETS A BALANCE SHEET

At first glance, these developments belong in different sections of the newspaper.

Gold and Bitcoin are trading stories. Treasury yields belong to macro. Data-centre finance belongs to credit. API pricing belongs to technology.

The connection is the increasing importance of the claim behind the asset.

A Treasury security depends on future sovereign payment capacity. A long-term data-centre lease depends on future cash flow. A purchase commitment assumes future demand. Debt financing turns that demand assumption into an obligation before the revenue necessarily arrives.

The assumption of continuously cheaper inference is different because nobody has signed it. Enterprises can still build capital plans and operating models around it.

Gold, silver and Bitcoin sit outside most of those arrangements. That does not make their prices stable or their valuations self-evident. It means they carry a different type of risk.

They do not require a borrower, customer or government to deliver a future cash flow.

That feature appears to be becoming more valuable at the same time that the market is becoming more interested in the balance sheets supporting everything else.

MANAGEMENT MARKS ITS HOMEWORK

The external conditions ledger exists because a forecast that cannot fail is not much use.¹⁶

The conditions filed previously therefore matter more than whatever confidence management attached to them at the time.

EC-001 — The long bond still demands additional margin

Condition: 30-year Treasury yield remains at or above 5.00%.

Status: SUSTAINED.

The long end remained above the threshold and traded at its highest yield since 2007 during the week.

EC-002 — The long end remains separated from the 10-year

Condition: 30Y–10Y spread remains at or above 40 basis points.

Status: SUSTAINED.

The curve continues to attach a material premium to additional duration.

EC-003 — Weak employment can coexist with expensive long-term money

Condition: 30Y–2Y spread remains at or above the filed threshold.

Status: SUSTAINED.

The filed relationship remains intact. The newer services survey points to firmer activity and hiring, which makes the explanation for elevated long-duration borrowing costs less one-sided than it appeared when the condition was first recorded.

EC-004 — France retains a material premium to Germany

Status: SUSTAINED / NEXT MONTHLY OBSERVATION PENDING.

No evidence presently requires withdrawal of the claim.

EC-005 — Labour-market weakening persists

Status: SUSTAINED / NEXT RELEASE PENDING.

The filed unemployment condition remains intact, although the August services survey suggests hiring momentum may be improving. The next official employment release will test whether that improvement appears in the underlying labour data.

EC-006 — The long end is pricing more than the economic cycle

Condition: 10-year Treasury yield remains at or above 4.50%.

Status: SUSTAINED.

The ten-year remains above the filed threshold. Stronger services activity may explain part of the level, but fiscal supply, inflation uncertainty and the cost of duration have not disappeared from the calculation.

EC-007 — The August long-bond auction was a receipt, not a ceiling

Condition: 30-year Treasury yield remains at or above the auction yield filed in the prior letter.

Status: SUSTAINED.

The subsequent expansion of Treasury buybacks gives this condition an additional policy test. The threshold continues to hold.

NEW FILING: EC-008

EC-008 — Alternative stores of value are expressing a common macro bid

Claim

Demand for non-sovereign stores of value is increasingly coexisting with elevated long-duration U.S. sovereign borrowing costs.

The purpose of this claim is not to declare Bitcoin equivalent to gold. It is to test whether their recent convergence represents something more durable than a temporary correlation.

Initial observation

The condition was filed on 22 August 2026, one day before this letter’s publication. Its original observation remains the measurement baseline.

At filing, gold, silver and Bitcoin have increasingly participated in the same broad move while the 30-year Treasury yield remains above 5%.

Failure condition

The claim will be considered breached if both gold and Bitcoin fall below their 22 August 2026 filing levels on a sustained basis while the 30-year Treasury yield remains above 5.00%.

Silver will remain a corroborating observation rather than a formal component of the failure test.

The filing-date levels will be preserved so the threshold cannot be improved retrospectively.

WHAT WE ARE WATCHING

The next round of evidence is already scheduled.

Jackson Hole should give us a better sense of how the Federal Reserve interprets stronger services activity, softer manufacturing and inflation that remains above target. The long end will show whether Treasury's larger buyback programme can produce a lasting change in the term premium rather than a temporary positioning move.¹⁷

Oil will determine how much of July's inflation relief survives into the next set of data.

The AI financing market deserves similar attention. The relevant question is whether the next increment of infrastructure continues to be funded primarily by cash-rich corporate balance sheets or whether debt, guarantees and externally financed vehicles take a larger share.

On the demand side, enterprise AI spending will eventually reveal whether the adoption forecasts survive more normalized model economics. DeepSeek's pricing change and the reported increase in AI-server costs are individual observations rather than evidence of an industry-wide reversal. They are still useful to watch when so much of the infrastructure being financed today assumes extraordinary amounts of future inference.

None of these observations requires management to know the answer in advance.

The point of the ledger is to establish what would change our mind before the market does it for us.

For now, the balance sheets are becoming more interesting than the promises attached to them.

Management will continue reviewing the collateral.

Market observations are timestamped where trading hours differ. Treasury buybacks are not Federal Reserve quantitative easing. Future corporate commitments are not uniformly debt, and the EC-008 failure condition excludes silver.