Read more

1) KEY METRICS / DATA POINTS

The Treasury announcement Wednesday morning is where this week starts. Secretary Bessent said the Treasury would at least double the size of its long-end buybacks, from $2 billion to $4 billion per operation, starting September 9 and running through November 4. The move targets the 10 to 20-year and 20 to 30-year sectors. What made this unusual was the timing. Treasury had published its quarterly buyback schedule just two weeks earlier without any mention of this. Mid-quarter surprises from Treasury do not really happen often. And when they do, it means something has changed.

Of course, the trigger was Tuesday’s print. The 30yr yield touched 5.34%, which is the highest reading since June 2007 as said before. We also said that for most of the past two decades, anything above 4.5% on the long bond was considered high. Sitting at 5.30% is a different world entirely. This level is being driven by a combination of things: sticky inflation, a $2 trillion annual deficit, uncertainty about where Fed policy is headed, Japan’s own rising yields making US long-duration less attractive to foreign buyers, and a growing sense that the term premium investors demand to hold long US debt is no longer what it used to be.

The buyback announcement produced a sharp initial reaction. The 30yr fell 9 basis points to 5.196% on Wednesday and the 10-year dropped to 4.647%. At the same time, stocks rallied. Then the trade fell apart. By Thursday, yields had climbed back to where they were before the announcement. By Friday, the 30yr finished at 5.273%, which was higher than where it started the week. Bessent followed up on CNBC saying he has “a big tool kit” and that the buybacks could go beyond $4B. The market read that as an admission that the pressure is not going away soon.

The July FOMC minutes came out Wednesday afternoon as well. As said before, 3 officials formally dissented at the July meeting in favor of a hike. What the minutes revealed is that this hawkish sentiment inside the committee is much bigger than the three votes. The minutes said “many participants assessed that policy tightening would likely be necessary if inflation did not decline.” 17/18 officials saw inflation risks tilted to the upside. The minutes specifically flagged the Middle East conflict as a concern for inflation, and pointed to elevated inflation expectations given inflation has been above the 2% target since 2021.

One important detail is that Warsh raised the possibility of reducing FOMC meetings from 8 per year to 6. He asked for input, made no decisions, and confirmed no changes would affect 2026. This all tells us how Warsh wants to run the Fed. Fewer meetings mean more accumulated data between them, which fits with his general instinct to say less and let the data speak for itself.

The 2yr Treasury yield finished the week near 3.85%, still reflecting the market’s growing conviction that the Fed will cut this year. That leaves the curve genuinely steep, with a 2s10s spread near 80 basis points. It shows us that the front end of the bond market is pricing cuts. The long end is not participating in that story. The bond market is voting that even if the Fed eases, structural pressures on the long end will keep it elevated.

Jackson Hole starts coming Thursday. Warsh speaks Friday at 10 a.m. Eastern. He has publicly framed the speech as a “blank piece of paper” and said he plans to focus on structural questions rather than near-term policy signals. Roughly 69% of fund managers in a Bank of America survey expect his tone to be neutral. This is the most anticipated Fed communication event of the year, but expectations for a market-moving speech are actually low, which itself is a setup where a surprise in either direction could produce an outsized reaction in the markets.

Oil finished the week at Brent $94.39 and WTI $87.06, both up around 5% on Bessent’s Thursday comments about “the toughest sanctions in history” on Iran. Gas at the US pump is at $4.11 per gallon, up nearly $1 from a year ago. The US Navy continues to run a covert convoy operation through the strait’s southern channel, keeping about 10 million barrels a day moving out despite the diplomatic breakdown.

2) AAIG’S PERSPECTIVE

I will tie this all together again.

So, the Treasury Secretary is stepping into the bond market to try to lower long-end yields. He is doing this because those yields have moved to levels that create real problems for the government’s own funding costs, for the housing market, and for corporate refinancing waves that hit in 2027 and 2028. He is also doing it because he cannot get the Fed to cut rates, and he cannot get the market to lower long yields on its own. So he is using Treasury’s own balance sheet to do what neither the Fed nor the market will do voluntarily.

This puts the Treasury Secretary and the Fed Chair in direct institutional tension. Warsh is running policy tight because inflation is still elevated. Bessent is trying to loosen financial conditions on the long end because fiscal costs are getting uncomfortable. When these two things are working in opposite directions, one of them eventually has to give. Historically, the Fed has been the institution that maintains independence. Bessent’s actions this week are testing that setup in a way we have not seen in a long time. He has not directly attacked Fed independence, but he is functionally substituting for what a Fed cut would do, and he is doing it in a way that draws attention to the fact that the Fed is not cutting.

A $2B to $4B increase in buyback size should be enough to push the 30YR down enough. However, the fact that yields unwound the entire rally within 24 hours tells you that the bond market does not think the tools Bessent has admitted to using are sufficient to change the trajectory of long yields.

There is a deeper issue here that goes beyond the immediate mechanics. If Treasury has to keep escalating its interventions to hold long yields down, and if the Fed refuses to help by cutting, we end up in a scenario where the Treasury Secretary is increasingly seen as the effective central banker of the long end while the Fed manages only the front end. This is a version of institutional split that has real consequences for how markets price US risk. It is the kind of thing that shows up in the dollar over time, in gold prices, in foreign demand for Treasuries. None of these are moving dramatically yet, but the setup for them to start moving is now in place.

The FOMC minutes complicated Warsh’s position further. Going into this week, the story was that Warsh had inherited a committee where three dissenters wanted to hike. Coming out of the minutes, the story is that a large majority of the committee sees inflation risks tilted higher, and many of them think tightening will be needed if inflation does not decline. That is a much harder committee to lead if you want to signal a dovish shift. If Warsh at Jackson Hole tries to open the door to cuts, he is doing it against the wishes of most of his committee. If he holds the line on data dependence, he leaves the September meeting genuinely open, but he also implicitly sides with the hawks against Bessent’s push for lower rates.

This is where Jackson Hole becomes interesting. The speech is normally about signalling policy direction. Warsh has said he does not want to signal policy direction. What he actually needs to do is signal something about Fed independence. If he says nothing about the Treasury intervention, the buyback program is validated as a legitimate tool for the executive branch to use in managing yields. If he addresses it, he is picking a fight with his own administration in his first Jackson Hole speech, which is not what a new chair typically does. The one he chooses will tell us a lot about how the Warsh Fed is going to operate.

One more thing on the yield story. The traditional read is that steepening reflects expectations of Fed cuts, which is bullish. That is partly true here. But the more interesting part is that the long end is not rallying with the front end. Historically, when the Fed cuts, all yields fall, mortgage rates come down, and housing rebounds. If the 30YR stays at 5.0%+ through a Fed easing cycle, that traditional relationship breaks, and housing does not get the boost. Corporate refinancing does not get cheaper. The interest-sensitive parts of the real economy see very little of what Fed cuts are supposed to deliver, and the fiscal problem gets worse because government interest expense stays high. That is a different environment than the last few easing cycles. Markets are continueing to assume that lower short rates flow through to lower discount rates on long-duration cash flows.

3) WHAT TO WATCH

First, of course the Warsh’s Jackson Hole speech Friday at 10 a.m. Eastern. Watch two specific things beyond the usual language on labour and inflation. One, does he mention the long end at all? Does he mention term premium, foreign demand, or the Treasury market? These would be unusual topics for a Fed chair speech, and any of them would signal that the Fed sees the yield situation as part of its own responsibility. Silence tells you the opposite, that the Fed views the long end as Treasury’s problem. Two, does he say anything about Fed independence, either directly or through framing? If Warsh addresses the tension, he is drawing a line. If he ignores it, he is choosing not to fight this battle publicly.

Second, whether the Treasury buyback program holds long yields at or below 5.20% next week. If yields keep drifting at this high level, we should expect Bessent to announce additional measures. If yields stabilise, the market is giving him the benefit of the doubt. If yields keep rising, the market is telling him he does not have the tools to solve this alone.

Third, the August jobs report on September 5 and the August CPI on September 11. These prints determine the September FOMC decision. We should watch how the bond markets react to the printed numbers. If weak jobs plus soft CPI moves the front end down another 25 basis points but the long end stays flat, you have confirmation that the yield curve is going to steepen structurally through this cycle. If both the front end and long end move down together, you get the traditional relationship back and the equity market can breathe. If the long end actually moves up on cut expectations, that is a genuine warning sign that the market is worried about fiscal dominance.

Thanks for reading this pulse of week 34, and see you at the next one!

AAIG is a research collective built on deep work, shared scrutiny, and quality over quantity. We treat research as a product, not content to fill a feed. By combining the expertise of multiple specialists, we find true edge where others just see noise. No hype, no daily trades. Just high-conviction ideas.

This is Asset Alliance Investment Group: Learn more on substack.com/@aaig

Disclaimer: The research and analysis published by Asset Alliance Investment Group is for informational purposes only and does not constitute financial, investment, or professional advice of any kind. All investment decisions carry risk, and recipients are strongly encouraged to seek indepen

The thing I want to spend most time on though, is the 30-year Treasury yield. It’s sitting above 5%, which is a level we haven’t seen sustained since before the 2008 financial crisis. It matters for almost everything else we’re going to talk about. Let’s get into it.

1) KEY METRICS / DATA POINTS

Let’s start with Wednesday’s CPI numbers. Headline came in at 0.1% for the month, bringing the annual rate to 3.4% from 3.5% in June. Core rose 0.2%, with annual core easing from 2.6% to 2.5%. Every single number matched what economists were expecting. Energy fell modestly on the month, though it’s still up 17.6% from a year ago. Shelter did most of the work in keeping the number positive and eased slightly to 3.2% annually. Airline fares picked up 2.2%. Medical care and used cars each rose 0.4%.

Here’s the thing about a print that lands exactly on consensus: it doesn’t settle the debate. If it had come in hot, the three hawks at the July FOMC meeting who wanted to hike would have been vindicated. If it had come in soft, the doves would have won September outright. Landing right in the middle means everything now depends on the next batch of data. Markets responded by trimming the odds of a September hike to 42% and pushing cut odds higher. That said, real wages are still negative for the fourth month running. Middle- and lower-income households are absorbing the squeeze.

Now the Treasury market, because this is where things got interesting. The 2-year yield dropped sharply on Wednesday and closed the week around 3.85%. That tells you the market is increasingly convinced the Fed will cut this year. The 10-year finished near 4.30%, well down from the 4.55% peak we saw in early July. But the 30-year is the one to focus on. It stayed above 5.05% all week and hasn’t really come down at all, even as the shorter maturities have rallied. Let me put that in context. For most of the past two decades, anything above 4.5% on the 30-year was treated as unusual. We are now sitting at levels last seen before the financial crisis, and the market is not showing signs of pulling that number back down. We will discuss the effects of this in the next section.

The Iran picture came apart through the week and formally ended yesterday. The 60-day MOU from June 17 expired, and an extension is not reached yet. Both sides are still blaming each other. Iran’s Foreign Ministry said the 60-day window never really started because Washington broke the understanding within 48 hours. A senior Iranian official told Reuters that Tehran has now moved from a defensive to a fully offensive posture. Trump, sitting in the Oval Office on Monday, said he was in no rush to extend anything, and separately threatened to bomb Oman if it interfered with the Hormuz situation. That’s the second time this year he has publicly threatened Oman, which is worth pausing on because Oman is functionally the intermediary keeping the entire back-channel process alive.

Traffic through Hormuz fell 19.5% over the week. From 19 crossings on August 11 down to just three on August 16. Of the 95 total crossings during the reporting window, 51 used the route Iran established and 44 went another way. Two attacks on August 11 killed six seafarers and injured ten more. Oil moved, but not dramatically. WTI is trading around $84 and Brent near $87. Elevated, but nothing like the panic we saw in July.

The Pentagon also announced a $22.9 billion contract with Raytheon on Monday for Tomahawk missiles, described as an attempt to fix a shortfall caused by all the spending during the Iran war. That’s part of the defense buildout we’ve been flagging for weeks. It’s feeding into industrial capacity, metals, propellants, skilled labour, and it’s one of the reasons core inflation isn’t coming down as fast as headline.

Japan went quiet after the coordinated intervention on August 3. JGB yields held around 2.80% through the week. The yen strengthened a bit. No further intervention needed. That’s a short-term relief, but nothing has actually been resolved. Takaichi’s autumn fiscal package is still coming.

2) AAIG’S PERSPECTIVE

Let me tie all this together, because the individual pieces are less interesting than what happens when you put them next to each other.

Let’s start with the CPI print. It gives the Fed nothing to work with. September now depends entirely on the August jobs report on September 5 and the August CPI on September 11. That’s a lot to hang on two data points, and it’s why Jackson Hole next week matters more than usual. Warsh has to say something meaningful about how the Fed is reading this without knowing what those two prints will show. His options are limited. Lean hawkish and risk being wrong if the labour market keeps softening. Lean dovish and risk losing the three dissenters who wanted to hike in July. Or hold the line on data dependence. He has done option three at every opportunity so far. My guess is he does it again, which means September stays open and the volatility around the August data becomes even sharper.

Now the 30-year, because this is what I really want you to sit with. When the long bond stays above 5% for a sustained period, it has consequences that most investors are not fully pricing in. Let me walk through them.

The first one is government interest expense. Every basis point higher on the long end costs billions in additional debt service over the life of new issuance. The deficit is already running hot from defense spending and the AI-adjacent industrial policy layer. High long-end yields are quietly making a fiscal problem worse, and the market has been happy to look past it. That won’t last forever.

Second is housing. Mortgage rates are stuck above 7% because the 30-year mortgage is priced off long-end Treasuries with a spread. If the long end doesn’t come down, mortgage rates don’t come down. Which means existing homeowners with sub-4% mortgages have no reason to move, and new buyers can’t afford to buy. The housing market stays frozen. The Fed can cut short rates all it wants, and it doesn’t fix this.

Third is corporate refinancing. A lot of the debt companies took out during the pandemic-era ultra-low-rate window has to be refinanced in 2027 and 2028. Companies that borrowed at 2 to 3% are looking at 6 to 7% or more when they roll. That’s a real drag on future earnings and on capex plans, and it’s a wave that hasn’t hit yet.

Fourth is valuations for long-duration assets. Growth stocks, real estate, private equity portfolio marks, anything with cash flows far in the future gets discounted more heavily when long rates are elevated. The math is simple. Higher discount rates mean distant cash flows are worth less today. The continued strength of the equity market alongside these yield levels is unusual, and honestly, the debate about whether the AI trade holds is increasingly a debate about whether the long end can come down.

Here’s the biggest implication of all this. If the Fed does cut this year, those cuts won’t do what history suggests they should. The traditional playbook has cuts pulling all yields lower, mortgage rates falling, corporate borrowing costs easing, and housing rebounding. If the 30-year stays at 5% while the Fed takes the front end from 3.75% down to 3%, the interest-sensitive parts of the real economy get very little of the relief. The trades that benefit are short-dated Treasuries and floating-rate credit. The trades that don’t benefit are housing, long-dated growth equity, and commercial real estate. That’s a genuinely different setup than what we’ve seen in the last few easing cycles, and it changes how you should think about the value of a Fed cut.

Let me be direct about the September pricing. The market is at 42% odds of a hold, with the rest weighted toward a cut. Six weeks ago, the market was pricing near-zero odds of a cut. That’s a real repricing, and it happened almost entirely because of the weak July jobs report. Wednesday’s CPI didn’t extend it, but it didn’t reverse it either. The market has decided the labour market weakness matters more than the still-elevated inflation. Warsh has not signaled agreement with that read. Whether he does at Jackson Hole is the question of the next 10 days.

On Iran, the market response to the MOU expiring is telling in its own way. Oil at $87 with a formally lapsed agreement, both sides accusing each other of bad faith, and Iran declaring a shift to fully offensive posture is not a market pricing danger. It’s a market that has learned that the pattern of the last six months (escalation, threats, private climbdown, quiet talks resume) is likely to continue no matter what any given headline says. That might be the right base case. It also means the market is thinly positioned for the scenario where the pattern breaks. Iran actually pursuing an offensive posture, or Trump actually striking Iranian infrastructure or Oman, would move oil quickly. Neither is my base case. Both are more plausible than the current implied volatility in energy markets suggests.

The Trump-Oman dynamic is worth thinking about carefully. Oman has been the diplomatic hub since April. It’s a Gulf state whose economy depends on Hormuz functioning, and it cannot afford a war. Threatening it publicly, twice in three months, is unusual behaviour toward the intermediary keeping the whole back-channel process alive. If Trump escalates against Oman, the process could collapse, and there’s no obvious replacement. Qatar is preoccupied with Gaza. Pakistan already delivered the June MOU and would face political costs at home for another round. Turkey has its own complications. The list of neutral parties trusted by both Washington and Tehran is short.

Let me zoom out on the Fed for a moment. Growth is decelerating. The labour market is contracting once you correct for the shrinking labour force. Inflation is coming down, but it’s still well above target. Real wages are negative. Consumer spending has been carried by higher-income households and credit for months now. Corporate earnings look strong on the surface at 37.9% growth in Q2, but that number is heavily influenced by AI capex accounting and by defense and healthcare spend. The Fed’s rate is at 3.50-3.75%, and market pricing has swung from cuts to hikes to cuts again in the space of eight weeks. The Fed is trying to hold policy in a middle ground while the underlying economy shows real signs of strain. And the long end of the bond market is telling you it does not believe cuts will fix the inflation and fiscal picture.

Japan, finally. The intervention on August 3 did what it was designed to do. It bought time. It didn’t fix anything. Takaichi’s fiscal package is coming in the autumn. The BOJ hasn’t signalled further rate hikes. The 10-year JGB yield at 2.80% is still elevated compared to recent history, and the pressure on the long end of the US Treasury curve remains. This is a slow-burn issue. It doesn’t move markets on any given week. It becomes suddenly consequential when the underlying stress hits a threshold. That threshold gets tested this autumn. A second coordinated intervention would be the signal.

3) WHAT TO WATCH

Three things over the next two weeks.

First, Jackson Hole on August 27-29. Warsh’s first speech as chair at the annual symposium. This is the most consequential communication event before the September FOMC. Every word will be parsed. Watch for language on the labour market. If he acknowledges softening that warrants “adjustment,” the market prices a September cut and the 2-year yield drops further. If he holds the line on data dependence, September stays open. Also watch for any signal on the long end. If Warsh addresses concerns about the 30-year yield or foreign demand for Treasuries, that would be an unusual topic for a Fed chair speech, and it would confirm the long end is now part of the monetary policy calculus. If he doesn’t mention it at all, that tells you the situation is being managed through channels other than the Fed.

Second, what happens in the strait now that the MOU has expired. Base case is more of the same. Talks continue informally, oil stays in the $80 to $90 range, occasional escalation triggers short-lived price spikes that resolve within days. The scenarios that break this pattern are Trump actually striking Iranian infrastructure, Trump escalating against Oman, or Iran making good on the fully offensive posture with a sustained attack on Gulf shipping. None of these is my base case. All three are plausible enough to warrant attention. The specific indicators to watch are daily Hormuz transit counts, currently three per day versus a pre-war run rate of 129, and any change in Trump’s Truth Social tone on Oman and Iran.

Third, the August jobs report on September 5 and the August CPI on September 11. These two prints determine the September FOMC decision. Weak jobs plus soft CPI moves the Fed to cut. Firm jobs plus sticky core CPI holds September. Anything in between, which is what I expect, leaves the meeting genuinely uncertain. The bond market’s reaction to these prints will tell us whether the long end finally starts to move, or whether it stays anchored at these elevated levels. If the 30-year doesn’t come down even as the Fed cuts, the message for the rest of the year is clear. Monetary easing helps the front end and not much else. The real economy has to live with high long rates for longer.

The thread running through all of this is that every major variable is in motion at the same time. The Fed, Iran, Japan, oil, the labour market, the consumer, the shape of the yield curve. The interactions between them are more consequential than any single one. The market is pricing continued muddle-through. It has held for six months, and it might hold for another six. But it has more failure modes than the current implied volatility suggests, and those failure modes are not independent of each other. A weak jobs print combined with an Iran escalation combined with a Japanese fiscal expansion is not a set of events any one of which is priced as dangerous. Together, they would look very different.

That doesn’t mean disaster is coming. It means the range of plausible outcomes is wider than the current price of risk assets implies, and the tails are correlated in ways that are worth thinking about. Position accordingly.

Thanks for reading this pulse of week 33, and see you at the next one!

AAIG is a research collective built on deep work, shared scrutiny, and quality over quantity. We treat research as a product, not content to fill a feed. By combining the expertise of multiple specialists, we find true edge where others just see noise. No hype, no daily trades. Just high-conviction ideas.

This is Asset Alliance Investment Group: Learn more on substack.com/@aaig

Disclaimer: The research and analysis published by Asset Alliance Investment Group is for informational purposes only and does not constitute financial, investment, or professional advice of any kind. All investment decisions carry risk, and recipients are strongly encouraged to seek independent, qualified financial advice before acting on any information contained herein.

In this article, we will walk you through our findings, where the company stands today and, more importantly, where the underlying market is heading. Our conviction in the long-term potential of two-phase liquid cooling has not changed. But the market opportunity timing and the company-level execution story are two different questions, and right now, Innventure and Accelsius face some very real challenges.

Accelsius is still working on a technology that solves a real problem. This is strenghtend by recent market commentary of the major players. As AI racks move to higher densities, existing cooling systems will eventually reach their limits. But adoption is taking longer than management expected, while Innventure is already operating with a tight balance sheet.

The stock fell 44% in after-hours trading and closed the following session down roughly 55% with another 14% drop the day after. To put it lightly, the market did not like what they saw. Management suspended its previously communicated 2026 revenue and cash flow expectations for Accelsius, not lowered it, suspended it. During our follow-up call, management acknowledged that the release could have been worded more clearly and did not fully capture the opportunities it is currently seeing. That provided some additional context, but it did not resolve the visibility issue: management did not offer revised guidance or any new quantitative outlook.The Q2 10-Q again contained a formal going-concern qualification.

The word again matters. The going-concern language was already present in Q1, so it was not a completely new warning introduced this quarter. But it also did not go away after the capital raised during the first half. That is a important part investors should focus on.

We still like the underlying technology as seen in our earlier articles. Two-phase direct-to-chip cooling solves a real problem as rack density increases. Accelsius has published technical work, demonstrated NeuCool on commercially available hardware and released independent benchmark results supporting its thermal claims.

The investment question at this point in time is not if whether the physics work. It is whether Innventure can finance Accelsius through a slower adoption curve and reach material commercial revenue on terms that still leave enough value for current shareholders. We will also be watching whether Accelsius can convert what management describes as its “leading position” in two-phase liquid cooling into revenue and reference designs. These are the key proof points we will follow closely, both for Accelsius and for the two-phase liquid cooling market as a whole.

The Adoption Delay

Management expected broader adoption of two-phase liquid cooling to arrive earlier. In practice, operators have continued extending the useful life of single-phase systems and existing cooling infrastructure.

Read more

as you know is focused on the retrofit model. The model gives WhiteFiber speed. It can take an existing industrial site, reuse part of the building and infrastructure, and bring compute online faster than somebody starting with an empty field.

If you combine that speed with reliability you have 2 major things the current market place is needy for. In both cases I view that management is taking the right first step. WhiteFiber is focusing on large multi-year customers before trying to maximize the price of every available megawatt. Of course they secure favourable pricing but the underlying focus is building further on reliability and showing that to customers and the market. Important step which you need to nurture at the right pace because it will pay off.

The contracts they are signing do three jobs. They fill the first deployments, give lenders something they can underwrite and build the operating record WhiteFiber still needs. If management executes on that base, the speed of the retrofit model becomes much more valuable later, because WhiteFiber can start choosing between colocation, its own cloud, managed services and potentially a smaller pool of short-term capacity. (What we saw in Nebius Q2, the 40-50 mil “burst customers”

WhiteFiber is first using long-term contracts to improve credibility and execution success based on key KPI’s the market is hungry for. That credibility should then create more freedom over how the next capacity is sold/marketed.

WhiteFiber management on what larger infrastructure buyers now prioritize.

Credibility comes before premium optionality

in terms of that part of the strategy a important comment on the call was that larger buyers are consolidating their deployments among a smaller group of providers. Pricing still matters, but engineering credibility, deployment execution and reliable ongoing operations are moving higher on the decision list.

That is logical. If an AI cluster is late or regularly misses its SLA, the customer does not only lose a few hours of compute. Its own model release, product roadmap or inference service can be delayed.

Michael Francisco (who was overall very strong on the call imo) made that point very clearly. More mature AI labs and enterprises understand the balance between price, quality and the total return on the cluster. Some customers have already been burned by cheaper providers through downtime and missed SLAs.

Michael Francisco on why customers increasingly value quality over the cheapest GPU price.

WhiteFiber starts to prove that its retrofit speed comes with reliable execution. They showed it with the Cerebras deployment and now with the NC-1 phase 1 part. The Cerebras deployment is running for a while now and is managed successfully. I do think WhiteFiber needs some more deployments and track record to really move it up a notch. but they are actively working on that. next to the right strategy and execution that simply needs slight a bit more time. Every successful deployment then reduces the risk assigned by the next customer. That is how the company can align proven company strategy and execution toward what fast, reliable capacity is actually worth.

The larger operating team at NC-1 is part of the same strategy. WhiteFiber expanded the team across facility operations, engineering and customer support.

Supplier relationships matter for exactly the same reason. WhiteFiber needs the GPUs, networking, cooling, switchgear and the rest of the system to arrive on one schedule. One missing component can delay the full cluster, as the switchgear issue at NC-1 already showed.

WhiteFiber has disclosed an NVIDIA Preferred Partner position and relationships with Supermicro, Dell, HPE and QCT. That supports access and integration, but I did not find evidence that the newly announced B300 and Vera Rubin systems were bought directly from NVIDIA. could be trough Krambu also. (New partnership I will address later) The contracts show access to new architectures. Delivery and uptime will show what those supply relationships are worth.

Shortly on the CFO change; I do not see it as a issue. Justin Zhu was already WhiteFiber’s Senior Vice President of Finance and Chief Accounting Officer and has been with the business since its formation. This looks like continuity, not a financial reset, which is what I would want while the company is working through several project-level financings.

NC-1 shows the logic behind the retrofit model

NC-1 is the first real test of the strategy, and management explained well that they purposely look for sites with immediate power access but also expansion potential. This is something I really like and shows the thinking behind making this business model successful over a long time frame. Being able to combine time to power mixed with expansion capability is such a sweet spot.

Approximately 20 MW of IT capacity was available for customer installation and testing at the time of the call, and WhiteFiber expected the first 40 MW to reach full billing around the end of August after resolving the switchgear issue.

the same property also carries a much larger power option. Duke Energy is expected to provide the delivery schedule for another 45 MW of gross capacity, while a further 200 MW is being evaluated. If everything is eventually approved and delivered, NC-1 could reach approximately 300 MW gross. Management pointed out that this is not a short term target and that another site acquisition would be more probable in terms or getting MW online.

Do not model 300 MW today. Power under evaluation is not contracted power. Contracted power is not connected power. And connected power is not revenue until the facility, clusters and customer onboarding are ready.

The strategic point is that WhiteFiber did not have to buy a perfect 300 MW campus on day one. It bought a functioning first project with the possibility of becoming a much larger campus later. If the additional power arrives, the company can expand infrastructure it already controls and bring optionality there.

Management on the potential 300 MW path at NC-1 and the two deployment paths for MTL-2.

The next 45 MW will also test management’s commercial discipline. Nscale has priority notification under the current agreement, but WhiteFiber said it has received strong demand from other counterparties as well.

This is where not allocating the capacity too early has a edge ofc. Management can compare the customer, price, duration, prepayment, financing support and expansion potential. And will use that edge also to see what fits them the best at this time. The highest visible price is not automatically the best contract. A slightly lower price can create a much better return on WhiteFiber’s equity if the customer prepays more, has a stronger credit profile or supports cheaper project financing. Also in terms of credibility as I pointed out before there could be choices made that around certain customers with high demands and respectable name that makes it strategic more interesting to go that route.

Another strength that I want to point out is a widely trend adopted by neoclouds currently and that’s local community engagement. More datacenter builders are talking to communities because local pushback has become a real execution risk. Water usage, noise, grid costs and tax incentives can all slow a project down. If local opposition delays a retrofit, part of the time-to-power advantage disappears.

But WhiteFiber has a edge here. they held a community day and explained that NC-1 should use around 85% less water and produce less noise than the previous tenant. Reusing an existing industrial site does not remove every objection, but it gives management a stronger starting point than placing a new campus on undeveloped land.

Community support is now part of time to power.

WhiteFiber on community engagement and why retrofits can reduce local pushback.

MTL-2 is not indecision. It is capital allocation

Then on MTL-2. Which shows more on WhiteFibers thinking. It may only represent around 5 MW gross, but it shows how management wants to allocate scarce power.

WhiteFiber has kept two deployment models open. It can use the site for traditional colocation, where the customer provides more of the equipment and WhiteFiber monetizes the building and power. Or it can vertically integrate the site through its own cloud, which produces more revenue but also requires more capital, financing and operating execution.

I like that management has not automatically selected the option with the largest reported revenue.

And Michael Francisco pointed out something here what i really like. The cloud team has to earn access to WhiteFiber’s data center capacity. It needs to bring a customer and an economic case strong enough to displace the colocation demand already available to the data center team. It was being talked about in the call by CEO Sam Tabor and Michael if it would make sense to join teams together, the data center team and the cloud team. And as it’s being discussed internally and it has advantages and disadvantages, I really like that Michael pointed out that keeping them separate keeps them on the edge in terms of execution. Don’t take the easy path, make them earn it. I like it. This is very healthy thinking if you want to be a top performer.

Michael Francisco on why the cloud business has to earn access to WhiteFiber’s own data centers.

Keeping both paths open gives management time to compare real contracts. If the cloud customer produces a better return after equipment financing, prepayments and execution risk, WhiteFiber can vertically integrate. If cloud only looks better because the revenue number is larger while requiring much more equity, colocation can be the smarter deal.

The networking platform makes imperfect power more useful

When you connect NC-1 with MTL-1, MTL-2 and MTL-3, WhiteFiber’s site strategy again becomes clearer. Management is not searching for the largest greenfield campus. It is looking for existing buildings where power, acquisition price, location and retrofit speed create an attractive route to compute.

The weakness (for now) is fragmentation. WhiteFiber may find 5 MW in one building, 7 MW in another and a larger block somewhere else. Each can be attractive on its own, but too small for a customer that wants one very large deployment.

The cross-data-center initiative is meant to reduce that problem. WhiteFiber demonstrated 111.2 Tbps of bandwidth with guaranteed sub-millisecond latency across 83 kilometers.

Michael Francisco on connecting separate sites and monetizing fragmented power.

If the platform works on real workloads, a smaller retrofit no longer has to carry the entire customer deployment by itself. WhiteFiber can remain selective on building price and speed to power, while combining several useful blocks into one logical environment.

That is what connects the Canadian sites to NC-1 strategically. The networking layer could make the site-acquisition strategy scale across imperfect power blocks instead of forcing management to wait for one perfect campus.

There is also a possible external model. WhiteFiber may use the platform internally and eventually license it across third-party facilities. That would extend the operating layer without requiring WhiteFiber to own every underlying site.

But this still has to be proven. Bandwidth and latency do not automatically make separate facilities equivalent to one physical supercluster. Storage, collective communication, orchestration and failure domains all matter. Management used demanding training workloads in the test, which is the right place to start, but I would not model several small sites as one large commercial cluster until a customer proves it.

For now, the value is simple: WhiteFiber is trying to make smaller power blocks worth more together than they are separately. And increase optionality and also agility. Something the market sees as favourable.

The contracts are financing credibility

Then on the new contract: WhiteFiber signed more than $540 million of aggregate contract value across new multi-year cloud agreements after the previous call. I care less about the headline than about what each contract proves.

Baseten brings approximately $165 million over three years and 1,392 NVIDIA B300 GPUs at a third-party site in Ontario. It is focused on production inference. That brings another reference case where WhiteFiber can show support live customer workloads where reliability matters every day.

Prime Intellect brings approximately $108 million over three years and 576 NVIDIA Vera Rubin 200 GPUs in Canada. This is WhiteFiber’s first disclosed Vera Rubin deployment and tests the platform on large-scale training and distributed compute using a next-generation architecture.

The relationship is also useful evidence. Prime Intellect started with smaller deployments before giving WhiteFiber the larger Vera Rubin agreement.

The Iceland agreement adds 576 B300 GPUs under a five-year contract worth approximately $87.5 million, with revenue-share potential. The Paris project adds a five-year agreement worth more than $160 million with an investment-grade customer.

Together, the contracts test inference, training, new architectures, different regions and different customer profiles. More importantly, they give WhiteFiber the operating history that the next large buyer and the next lender will want to see.

Michael Francisco also explained why customers are moving toward longer duration. H100 capacity can cost more per GPU hour today than when it entered the market. A customer that expects to need the compute for years is not only securing capacity. It is protecting itself from having to return to a constrained market later at a worse price.

Michael Francisco on why customers are re-evaluating longer cloud contract durations.

Burst demand is a future option in my opinion.

Management confirmed of course that they see power-ready capacity in 2027 as extremely scarce. Broader consensus for people who focus on that market. WhiteFiber’s retrofit model is built around reaching the market faster than a greenfield development. That speed becomes especially valuable when a frontier lab needs a large cluster fast and cannot wait for a normal construction timeline. Yet burst demand (the 40-50 per mw that Nebius is getting) is immediate capacity. WhiteFiber can‘t deliver that now but in my personal opinion their strategy could lead to a very agile company full with optionality. The smaller clusters and interconnect between DCs should support that.

The current multi-year strategy is what can make that opportunity real later. First prove engineering credibility, deployment execution and reliable operations. Then a customer can trust WhiteFiber with a critical training sprint. If management can also keep a small amount of capacity flexible, the retrofit speed may then deserve a completely different price.

That is why the open choices across MTL-2, third-party sites, networking and managed services matter. WhiteFiber does not have to speculate on burst demand today. It only has to avoid locking every future megawatt into one model before the market shows what that capacity is worth.

Managed services changes who funds the asset

Managed services may be the cleanest extension of WhiteFiber’s model. The customer funds the hardware and data center capacity. WhiteFiber deploys and operates the infrastructure through the systems, software and people it already has.

Michael Francisco explained the economic difference. WhiteFiber can earn margin from day one because it is not first working through the repayment of hardware financing or data center costs. Management expects the margin profile to move closer to software when additional service layers are included.

Michael Francisco on the capital-light economics of managed services.

This separates the value of WhiteFiber’s technical platform from ownership of the asset. WhiteFiber cannot fund every GPU and every attractive retrofit itself, but it can apply the same engineering and operating capabilities across infrastructure funded by customers, lenders or other operators.

That is strategically important. The same operating team can produce revenue across more infrastructure than WhiteFiber could ever own on its current balance sheet.

The Krambu partnership follows that logic. WhiteFiber receives exclusive access to up to 100 MW of liquid-cooled colocation capacity beginning in 2027. More interestingly, the companies are trying to solve GPU availability and power availability as one customer problem.

Customers often find the GPUs and the powered site on different timelines. WhiteFiber wants to combine both supply pipelines and give the customer one counterpart that can say what will be delivered and when. Management also said the partnership combines WhiteFiber’s OEM and NVIDIA relationships with Krambu’s own connections, which could improve access to allocations.

Michael Francisco on why Krambu combines the power and GPU timelines behind one customer offering.

I still would not treat the full 100 MW as delivered capacity. Krambu has to execute and WhiteFiber has to convert the agreement into customer deployments. But the strategic choice makes sense. WhiteFiber is not only reserving third-party space. It tries combining two lets say incomplete pipelines into one product that is easier for the customer to buy. Makes adoption and aligning workstreams much easier and efficient.

Project financing decides how far the model can scale

Then: WhiteFibers whole strategy becomes much more powerful if each contracted project can increasingly finance itself.

if you describe the goal you would want: A multi-year customer signs a firm contract and prepaid part of the service. Then an equipment lender finances part of the GPUs. A project lender lends against the infrastructure and contracted cash flow. WhiteFiber contributes less equity and can move that capital into the next retrofit.

The customer contract is therefore not only future revenue. It becomes part of the financing package. The prepayment reduces the capital WhiteFiber has to raise. Project-level financing is what prevents every new site from requiring another large corporate equity issue.

currently WhiteFiber used expensive bridge financing while working toward permanent project financing. The bridge carries a 9.5% contractual interest rate, but the filing showed a much higher effective annualized cost because of the minimum-return requirement and short duration. The bridge was meant to be temporary but is taking longer then expected and that is a risk.

Management is now in exclusivity with a lender consortium and working through diligence and definitive documents. If permanent financing closes on acceptable terms, WhiteFiber can repay the bridge, release part of the equity tied up in NC-1 and recycle it into the next project.

Management on recycling NC-1 capital into the next retrofit project.

The first refinancing would also create their wanted template. making the next lender underwrite a completed retrofit, a contracted customer and an operating record.

That is why the financing is both the largest near-term risk and an important catalyst. If it does not close, WhiteFiber remains dependent on expensive short-term capital and the next site becomes harder to advance. If it closes, the first full turn of the model is validated: acquire advantaged power, secure the customer, stabilize the project, refinance it and recycle the equity.

WhiteFiber ended the quarter with $56.1 million of cash and around $143 million of deferred revenue, mainly customer prepayments. It also added approximately $83.2 million of project-level equipment and bridge financing during the quarter.

That customer cash is useful, but it is not mature free cash flow. It funds deployments WhiteFiber still has to deliver. This is why I care much more about the quality, prepayment and financeability of the contracts than the quarterly revenue number.

The strategic takeaway

So let me extrapolate all of this into more concrete terms.

In my opinion, the retrofit model is strategically very strong, but only as part of a broader infrastructure platform. Of course, it will not replace greenfield because it solves a different problem.

I think that if WhiteFiber executes successfully through 2030, it does not necessarily have to be a smaller version of the other neoclouds. Maybe in portfolio size, but not in importance. I think WhiteFiber can become an operator that finds underused industrial power and converts it.

Retrofit should remain valuable throughout the coming years because power scarcity is unlikely to disappear simply because more projects are announced. At the same time, more projects are being put on hold because of different forms of pushback and potential regulation.

The bottleneck is also not only the electricity grid. It includes transmission, substations, transformers and everything else you can name. That supports WhiteFiber’s basic strategy. If you have an existing industrial property with an existing grid connection, it can have far more strategic value than the building itself.

WhiteFiber is essentially trying to buy that time.

WhiteFiber should also not see itself as real competition to greenfield because a purpose-built campus can support much larger contiguous power blocks, have much better electrical and cooling infrastructure and operate far more efficiently because of the building structure.

The larger frontier labs that want tens of thousands of GPUs operating as one cluster will also not operate at the same efficiency if you connect multiple locations through fiber. You can distribute some workloads across locations, but communication-heavy frontier training still benefits enormously from being in close physical proximity.

But I think WhiteFiber is very aware of that.

If you look toward 2030, I think the market will become more segmented. The largest frontier training clusters and inference loads will move toward gigawatt-scale, purpose-built campuses. But there will also be substantial demand for regional and sovereign AI capacity, production inference, enterprise training and fine-tuning, for example.

There will also be customers that need capacity before a large greenfield campus can be completed. I think that is a certain edge WhiteFiber does not promote as loudly, but it is something I want to go deeper into strategically when I speak with WhiteFiber.

Because if I am correct, the market is underestimating what this strategy can actually lead to.

That being said, WhiteFiber is of course not the only company focusing on an asset-light partnership model alongside its core business. The winning model across the market will probably be some kind of hybrid between greenfield, retrofit and asset-light capacity.

That validates WhiteFiber’s direction, but it also means that retrofit itself is not enough of a differentiation. Many companies can buy an old industrial building. Of course, WhiteFiber has more experience with it and can do it a little faster, but that is still at the margin. It is a skill that can be learned or bought.

One of the biggest things that makes the model more challenging, and that I see as a risk, is that WhiteFiber cannot use one complete reference design. Every site it buys will have a different layout, different constraints and different needs. Each site requires a custom design.

If WhiteFiber can standardize as much of its internal process as possible, including site selection, engineering and procurement, that will already provide a huge advantage and help make the model repeatable.

It is also why I pointed out that I like management keeping the data center and cloud teams separate. It prevents the cloud business from automatically being assigned every site or megawatt WhiteFiber controls. They need to earn that capacity.

All in all, I think WhiteFiber’s edge will not come from retrofitting buildings alone. It needs to combine certain aspects of its strategy.

I see it as finding power positions that others overlook, converting them faster through internal optimization and experience, financing each completed project at the project level, operating the infrastructure reliably and then applying the cloud and managed-services platform across both owned and third-party capacity.

If WhiteFiber can do that successfully through 2030, I think it can become a very interesting AI infrastructure operator with a valuable combination of owned power, cloud capability and asset-light operations.

That is a real strategic position if you build it over time. But the value is created by the execution, the strategy and the capital recycling, not by the word retrofit itself.

Nebius explained three types of contracts on the call. Long-term contracts with investment-grade customers help finance the buildout. You use these for acceptable terms to fund your operations and postion yourself for the most faborable client mix. These are the 5 year AI cloud agreements with the Hyperscalers. Second you have the one-to-three year contracts you focus on with enterprise, souvereign and frontier labs against a higher price. These are companies that provide a long run way of early AI adoption and are your best target client to build a customer trust connection with and commit together to scale in the future. I call them the high potentials. And then there is a smaller pool of capacity for customers that need a very large amount of compute immediately and only for a limited time. Those customers are willing to pay a very high price for it. Lets call them burst buyers.

And the interesting part is that Nebius could already sell all of its planned 2027 capacity today on the normal one-to-three-year terms. Management is choosing not to. That decision tells you quite a lot. They are building capacity in advance, but deliberately not selling all of it in advance, because they want to keep the ability to decide later where that capacity creates the most value.

Nebius is optimizing across customer, price, prepayment, duration and deal size, not simply selling every megawatt to the highest headline bidder.

See how they have set this up? The big stable contracts finance and utilize the base. Arround 70% of the deals closed in Q2 included prepayments covering 50% to 60% of the associated CapEx. Then the high potentials serve you lock in for commited future growth as their runway is much longer and will grow in capacity hunger. The company then can then keep a smaller portion flexible and wait for opportunities that deserve a completely different price. Nebius is not optimizing one number. It is optimizing the customer, price, prepayment, duration, deal size and the possibility that the relationship becomes much bigger later. Next to be able to feed the burst buyers which has the high pricing. (The 40-50 mil per MW) I will elaborate a bit further in the article how they build that up really smart. Lets show some fundamental changes first that the market gave insight in.

pricing signal move from $12 million to above $20 million and peak above 40 mil for burst demand buyers.

he normal Q2 deals moved above $20 million per megawatt before the short-term premium deals moved above $40 million.

The $40 million-plus short-term price gets most of the attention, but I think the move underneath it is more important for the long-term model. Nebius shows a 2026 base of about $12 million of ACV per megawatt and Q2 deals above $20 million. That is not a small pricing improvement. It means the normal contracts are moving up strongly before we even talk about auctions or customers that need capacity immediately.

I would therefore see roughly $20 million per megawatt as the first real indication of where the pricing floor for new normal deals may be settling, but not as a fixed price list. Management again was very clear that it optimizes across customer type, price, prepayment, duration and deal size. A contract at, hypothetically, $18 million per megawatt can still be the better deal if the customer prepays more CapEx, signs for longer, has a stronger credit profile or creates a much larger path into inference and future capacity. See it as a payment to optimize client mix to keep executing on that agile strategy. Agility keeps this strategy alive.

The opposite is also true. Nebius may ask above $20 million if the contract is shorter, the customer wants more flexibility, the deployment is more complex or the company gives up too much future capacity. So the strategic point is not that every new contract should be modelled at exactly $20 million. It is that the centre of gravity has moved from about $12 million to above $20 million, while Nebius now has enough demand to become more selective about which economics it accepts. Again agile strategy. In a fast moving market the agile strategy captures the best economics in my opinion. Strong base with agility on top.

That is also why I would not use a single ACV-per-megawatt number to value the whole build. Two contracts can have the same headline ACV and still be very different for Nebius. One can finance most of the hardware upfront and bring a new strategic customer onto the platform. The other can tie up scarce capacity for years without enough flexibility. The first can easily be worth more, even if its reported ACV per megawatt is slightly lower. Strategic importance vs economics is what management weighs these deals on.
This is a proven strategy troughout succesfull companies in history. A decision metric I personally used for a decade on all major deals i had to make for the company I worked for.

Do not apply the peak price to every megawatt of course. X is full of these BS headlines already.

Peak pricing is literally peak pricing for the burst buyers. You can‘t model the $40 million to $50 million per megawatt number and apply it to every gigawatt Nebius is going to build, because that is simply not in line with reality. This is annualized pricing for short-term deals that normally run for three to six months. The normal one-to-three-year contracts signed this quarter were priced at $20 million to $25 million per megawatt.

That annualized part is important. It does not mean the customer pays the full $40 million to $50 million for a three-month engagement. The actual contract value is adjusted for the shorter duration. But it does show the much higher rate Nebius can charge when the customer needs a dedicated cluster immediately, because the value sits in the timing and certainty rather than in using that capacity for a full year. It is also a leading indicator and metric on the power scarcity and online compute. You could also debate as the rising indicator for ROI on AI. that‘s a other debate.

You should see the $40 million to $50 million as premium pricing for customers that need a large amount of compute immediately and only for a limited time. Management mentioned dedicated GB300 clusters for a large training run ahead of a model release or for a reinforcement-learning post-training sprint. If a frontier lab needs a few extra weeks or months before launch, waiting for new capacity is not an option.

To break it down a bit for the more average investor: Think of it like renting a Ferrari for a wedding weekend. You pay a very high daily price compared with what you would pay to lease a Ferrari long term. But the customer does not need that Ferrari all year. They only need it for that specific moment. And for the rental company, which is Nebius in this case, the economics can be very strong if it can rent the same car repeatedly.

Frontier labs need immediate, time-boxed capacity for large training runs ahead of model releases and post-training sprints.

According to management this is fully aligned with how the market is developing. Even to their suprise as they said. Management was positioning on agility but sometimes you can’t see the whole playbook so this landed in their advantage. You need some luck sometimes also right? Management sees that more and more frontier labs and AI-native companies are coming online, and model launches are becoming bigger and more competitive. So the market for immediately available capacity gets bigger with them. Nebius does not need to price every megawatt this way. It needs enough urgent customers moving through a smaller flexible pool with as little idle time as possible. If it can do that, the ROI on that pool can be very high.

The auction tells Nebius what capacity is worth today

As briefly touched earlier the auction sits next to this short-term strategy, but it is not the same deal. The auction was a separate go-to-market motion for free, unallocated capacity. It cleared 15% above the highest price Nebius had charged before and 20% above its Blackwell pipeline. The winning customer was happy with both the pricing process and the certainty that the compute would actually be available, and already said it wants to participate again.

Nebius is using a small part of its capacity to let the market tell it directly what scarce compute is worth.

This is a very smart way to get actual price discovery in a market where standard reference points, competitor pricing and analyst estimates are all over the place. Nebius said there are several buyers for every GPU, so instead of guessing what the capacity is worth, it just lets the market decide it. Just let them bid.

But again, management said it is using only a small portion of the overall capacity for auctions and other price-discovery initiatives. The goal is to learn what a scarce block of capacity is worth at that exact moment, use that information in future negotiations, and keep the majority of the business on more predictable contracts. Right in the 3-lane strategy we pointed out in the start.

Build capacity in advance, but do not sell all of it in advance

Nebius adressed the model very plainly on the call: “Our model was built to build capacity in advance, but not to sell it in advance.”

In a market where prices and customer needs are changing this quickly, selling every megawatt years ahead removes your ability to react. Nebius is accepting some utilization risk in exchange for option value. Very strong strategic play by management.

Nebius is deliberately preserving free capacity so it can react to the market instead of locking everything away years in advance.

Of course, this only works if the stable base is large enough. Unallocated capacity can also sit idle, and if pricing turns before the hardware is placed, Nebius carries that downside. But with long-term customer commitments, more than $9 billion of expected prepayments in 2026 and asset-backed against contracted cash flows, the company is trying to finance the base while keeping that smaller layer open. See the way they have balanced that? Stability underneath, agility on top.

The customer mix is also a financing strategy

This is where the three contract types come together financially. The long-term investment-grade contracts are as pointed out not there because Nebius wants to lock the whole platform away at the lowest price. They are there purely because they make the build financeable. Customer prepayments reduce the amount Nebius has to fund itself, and the contracted cash flows can then support asset-backed debt on top.

Roughly 70% of Q2 deals included prepayments, while contracted cash flows already supported Nebius’s first asset-backed debt facility.

Nebius expects more than $9 billion of customer prepayments in 2026. It also raised a $775 million facility against deployed GPU infrastructure and the contracted cash flows from an investment-grade customer, while saying it has more than $40 billion of additional customer commitments on similar terms. So every stable contract can do two jobs. It produces future revenue, but it also helps pay for the capacity before that revenue arrives.

This matters because building ahead normally creates a funding problem. You spend the money first and only find out later whether the customer arrives. Nebius is trying to solve that by financing the stable base through prepayments and contracted cash flows, while preserving some capacity that has not yet been allocated. That smaller open part can then be used for the higher-priced short-term deals, auctions or a new strategic customer that suddenly appears.

So when Nebius evaluates a customer, price is only one part of the return. The company also has to ask how much CapEx the customer funds, whether the contract supports cheaper debt, how much capacity it locks up, how strong the credit is and what that customer can become later. This is again why a slightly lower ACV can still create more value for Nebius than the highest visible price.

The third-party strategy is where it becomes really interesting

Potential partners have capacity and capital, but need Nebius to build the product and bring it to customers.

Nebius said it received dozens of inquiries from potential partners that have significant capacity and enough capital, but do not know how to build the full AI cloud product or how to sell it. This is a gap Nebius deliberatly wants to fill. The partner owns and finances the infrastructure. Nebius brings the architecture, reference design, full-stack software and access to customers.

Put more plainly, Nebius can take somebody else’s power, building and GPUs and turn them into Nebius capacity. The partner gets a much faster route into the AI cloud market. Nebius gets extra capacity on its platform without paying for the entire underlying asset itself.

And this is why the multi-tenant platform matters so much. If the same Nebius platform can run across owned data centers, colocated capacity and third-party capacity, management can treat those different sources as one broader pool. It can use long-term customers for the stable base, move the short customers through the flexible part, and add capacity from partners when the market gives it an attractive opportunity.

This is also where the Ferrari example what I pointed out before has double use. Nebius does not necessarily have to buy every Ferrari itself. A partner can own the car, while Nebius provides the platform, customer access and the learned knowledge (or monitored) of how to keep it rented. If Nebius can standardize that partner capacity and move enough short-term customers through it, it can earn on the software, operation and sale of the capacity while using much less of its own balance sheet.

And I love to give credit to Nebius here. I already wrote a while back, when discussing its partnerships, that using partners to scale a third-party multi-tenant platform would make a lot of sense. In my opinion, agility is everything in this market. Capacity takes a long time to build, while model companies can suddenly need an enormous cluster for only a few critical weeks or months. Nebius management has positioned the company between those two timelines really, really well.

It also changes the relationship between revenue growth and Nebius’s own CapEx. If all growth has to come from owned data centres, every new dollar of capacity requires another large investment from Nebius. Partner-owned capacity breaks part of that link. Nebius can still earn from running the platform, selling the capacity and adding software and services, while the partner funds more of the power, building and hardware underneath it.

That does not mean the economics will automatically be better. Nebius has to share part of the return with the partner, standardize different sites and prove that third-party capacity performs as reliably as its own. But if it works, the company can enter more regions, react faster and grow the platform without putting the same amount of balance-sheet capital behind every megawatt. That is why management calls it asset-light, and why the model can become much more important than a normal capacity partnership.

The short-term deal can become a long-term customer

One customer called it the best POC it had ever had, and the discussion is already moving into more capacity, inference, post-training and Vera Rubin.

This is the second part that I think people will underweight. The premium short-term opportunity is not only about making a very high return on spare capacity. It can also be the start of the customer relationship. Nebius can solve the customer’s most urgent problem, prove the platform on one of its most important workloads and then move with that customer into post-training, inference and the next generation of GPUs.

The four large wins this quarter were all competitive. Every customer already had an existing supplier, in some cases a hyperscaler, and Nebius still won on scale, performance, reliability, responsiveness and technical support. These were not inbound walk-ins either. They went through multiple engagement cycles and hands-on POCs. One of them told Nebius it was quite literally the best POC it had ever had. (Quoted by nebius tho i take these statements more lightly)

And now Nebius is already talking with several of those customers about post-training and inference, more capacity and Vera Rubin. So the expensive first rental can become a multi-year customer. That is a much more interesting economic model than simply filling spare capacity for a few months.

This is why I do not see short-term capacity as a completely separate business. It can function as the front door to the platform. The customer first comes because timing is critical and Nebius has the cluster available. Nebius then gets to prove its technology during the exact workload that matters most to that customer. If it performs well, the conversation naturally moves from one training sprint into post-training, inference, more regions and the next generation of GPUs.

The quote about the best POC they had ever had is evidence of that. It does not prove every trial becomes a billion-dollar contract, but it shows Nebius is not winning only because it happens to have spare GPUs. The platform, responsiveness, transparency and technical support are part of why the customer stays. That is what can turn one premium transaction into long-term customer value. Locking in that partner in that ecosystem.

Customer diversification is also becoming more strategic

The customer mix also tells you how Nebius is building the company. Reflection, Cohere and the new US AI lab deepen the relationship with AI-native customers and frontier-model demand. The large US quantitative trading firm adds a completely different vertical. And the wider pipeline now spans AI natives, new labs and enterprises, with multiple opportunities above $1 billion.

Those customers do different jobs for Nebius.

Of course, Nebius is not fully diversified yet. These are still very large contracts, so customer concentration remains real. But the mix is moving in the right direction, and more importantly, it is broadening across customers that have different strategic value to the platform.

That diversification also gives Nebius better information. Frontier labs tell the company which training architectures and cluster sizes are needed next. Enterprises show where inference and specialized models can become recurring workloads. A quantitative trading customer adds another type of high-performance demand with completely different timing. The more of those customers Nebius serves, the better it can decide what capacity to build, where to build it and how much of it should remain flexible.

And that is important for revenue growth. The first sale can be compute, but the relationship can broaden into inference through Token Factory, post-training, software and support. Management already said asset-light revenue and those higher-value services should become a larger part of the mix. So the longer-term opportunity is not only more megawatts at a higher ACV. It is more revenue and potentially more margin from every megawatt that sits on the platform.

Nebius already points to higher utilization, asset-light capacity and inference services as additional sources of revenue and margin.

Why Nebius is already building for 2027

Nebius could already sell its full planned 2027 capacity on current terms, but is deliberately keeping part of it open.

The demand signal for 2027 is therefore much stronger than a normal pipeline statement. Nebius can already sell the planned capacity at $20 million to $25 million per megawatt, the signed deals have an expected payback below two years, the pipeline contains several opportunities above $1 billion, and existing customers are already discussing Vera Rubin capacity.

So Nebius is raising the contracted-power target to 5 GW and wants to deploy more than 1 GW of new capacity per year from 2027.

I think management’s confidence rightfully on 2027 comes from having several demand signals at the same time. The company can already sell the capacity on today’s normal terms. Existing customers are discussing more capacity and Vera Rubin. The pipeline contains several opportunities above $1 billion. The first auction produced a higher market-clearing price. And the short-term training deals show that customers will pay much more when timing matters. One signal can be temporary.

The strategic decision is therefore not simply to build one gigawatt because demand is high today. It is to create several routes for placing that gigawatt when it arrives. Some capacity can go to the contracted base. Some can remain multi-tenant. Some can be sold for a short training sprint. Some can come from a partner instead of Nebius’s own balance sheet. And if pricing keeps moving, the auction gives the company another way to see what the market is actually willing to pay.

That being said, contracted power is not the same as active, revenue-producing capacity. Management explained that after power is connected, it still has to commission the data center, build the clusters, deploy the platform and onboard the customer. That takes several months, which means part of the 800 MW to 1 GW expected by year-end 2026 will only become active throughout the first half of 2027.

Connected power still needs commissioning, cluster installation, platform deployment and customer onboarding before revenue starts.

So execution is still the part I am watching. The third-party model is early, and I want to see actual partner megawatts, how quickly those sites are deployed, what Nebius earns from them and whether they perform as reliably as owned infrastructure. Same as I also do not expect the $40 million to $50 million pricing to become the normal price across the whole platform. It only needs to remain attractive on the smaller flexible pool.

I am also watching whether the normal ACV really holds around the new level. The move above $20 million per megawatt is very strong, but future contracts will keep moving around it depending on duration, prepayment, customer quality and strategic value. If Nebius can keep the normal deal economics near that level while growing the flexible and asset-light parts on top, then the 2027 opportunity becomes much larger without needing the unrealistic assumption that every gigawatt earns the peak short-term price.

The strategic takeaway

Nebius is not simply building more megawatts. It is building several ways to source, finance and sell the same underlying compute. Long-term customers finance and utilize the base. Mid-term AI cloud contracts build the core business. A smaller flexible pool serves customers that need immediate capacity and are willing to pay a premium. The auction shows Nebius what scarce capacity is worth at that exact moment. And third-party partners can add more supply without Nebius funding every underlying asset itself.

See the way this comes together? Nebius is trying to keep the stability of an infrastructure company while adding the agility of a marketplace. It can serve a frontier lab that needs a critical cluster before launch, charge a high price for that urgent capacity, prove the platform during the customer’s most important workload and potentially keep that customer for years of post-training, inference and future architectures.

If this works, the real advantage will not be bare-metal access or power alone. It will be the ability to take owned, colocated and partner capacity, standardize it through one platform and continuously decide where every part creates the most value. That’s the nebius ecosystem and let me spoil it that can become highly profitable.

That is why I think Nebius executed this strategy really, really well. The company is positioning itself between long construction timelines and very fast-changing AI demand, and it is building the flexibility to make money from both. Now it has to prove that the third-party platform can scale and that the 2027 capacity arrives on time. But strategically, the way Nebius has positioned itself is extremely strong.

Unlike the earlier whitepaper from a few weeks back, this isn't third-party research on Accelsius's product. It's Accelsius showing off its R&D pipeline, one step beyond what it currently sells. That shifts how you read it and this isn't someone else validating Accelsius, it's Accelsius telling you what comes after the cold plate era, where two-phase liquid cooling tech can go, and more importantly, where data center cooling designs might be headed.

The PFAS-free paper came out of a collaboration with SK Innovation, so it's nice to see something from that partnership, which we touched on in our initial coverage and deep dive on Innventure. If you're new to the story, you can check out our previous Technical Series with our view on the whitepaper—it has a lot of background.

What They Tested

Let me start with the Direct-to-Die paper. To get what they actually tested, you first need to know how a GPU is put together. Current liquid cooling, whether it’s single-phase or two-phase, never makes direct contact with the chip, because between the silicon and the coolant you’ve got a Thermal Interface Material, or TIM, plus a metal cold plate. That whole stack is a hard limit on every design; it can only improve so much. So the paper goes after the obvious next step: what happens if you strip out the TIM and the cold plate entirely, and just boil the refrigerant straight on the die?

You can only ask that question with a dielectric fluid. Spray water-glycol onto live silicon and your server is dead. The dielectric refrigerant doesn’t conduct, so contact is harmless. Remember this point, because it matters for the market impact section: at the die level, liquid cooling has left the chat (again). The setup is easier to understand than it sounds. Accelsius built a copper block the exact size of a B200 die (60 × 26 mm) with heaters inside. Call it a fake GPU, basically, that they can crank up to any power level they want. They pushed it to 3 kW, which is 2.5 times what a real Blackwell chip produces. So whatever passes this test has headroom for the next chip generations, not just today’s.

The coolant is Honeywell’s R515b, a refrigerant developed specifically for data center cooling: non-conductive, non-flammable, and with a low global warming potential. So refrigerants keep a sensitive point in any industry and regulations so it’s worth to keep ahead of the discussion and put things in perspective, that’s why we updated the table from the Innventure deepdive to include this R515b type.

What the table shows is that R515b is a blend of two components, an HFO and an HFC component and their atmospheric lifetimes are two worlds apart. The HFO part breaks down in weeks; the HFC part sticks around for decades. That is why the blend sits at GWP 293 instead of the near-zero figure of R1233zd(E), the fluid in their shipping CDUs.

The PFAS free refigerant paper shows a development of a R-4101A product which is a zeotropic blend of R-32 / R-152a / R-13I1 under ASHRAE Standard 34. More on this later.

Back to the test, so they tested three versions of the die surface:

• Fins on the die. Microscopic cooling fins sitting directly on the chip surface, with refrigerant boiling between them. This simulates a future where the heat sink is bonded straight onto the chip, or channels are etched into the silicon itself. No TIM, no cold plate.

• Jets on fins. The same finned surface, but now with a grid of twenty tiny nozzles (half a millimeter each) spraying refrigerant straight at it.

• A bare die. A smooth, polished surface with nothing on it at all, the stand-in for simply spraying a naked chip and hoping the boiling does the rest.

The Microsoft Angle

Accelsius is certainly not the only one working on this. Microsoft has been public since late 2025 with the Swiss startup Corintis, TSMC is testing bonded silicon lids with microchannels etched into them, and Bernstein names Microsoft and TSMC as the ones to watch on direct-to-die. So this is a race that includes some of the largest companies in the industry, and Accelsius is one participant in it.

What is interesting is that they arrive from different directions. Microsoft and Corintis etch the channels inside the chip and seal them, which means ordinary water can be used because it never touches the live circuit. Accelsius sprays dielectric refrigerant onto the die in the open and lets it boil, which only works because the fluid doesn’t conduct electricity. Both routes end up in the same place: no thermal interface material, no cold plate, coolant reaching the silicon itself.

That difference decides who has to move first. The sealed route is a packaging change, so the chipmaker or the foundry has to build it in. There is no retrofit path and no installed base to sell into. The open route can in principle be retrofitted at server level, which is the market Accelsius already sells to, although it still needs fins bonded onto the die, so they are not fully independent of the chipmaker either.

Also if the fluid in the sealed channels is water-based, there's no fluorinated coolant, so no PFAS exposure on the coolant axis. That is a structural advantage over two-phase, and it's the strongest form of the competitive threat to Accelsius.

The Results

At the B200’s rated 1,200 W, the finned flow-boiling setup (Test 1) hit a case-to-fluid thermal resistance of 0.007 K/W. Put that next to the numbers from our whitepaper analysis and you see the whole staircase: Dell’s stock single-phase plate at 0.0328, the NeuCool two-phase cold plate at 0.0171, direct-on-die at 0.007. In temperature terms: at Blackwell power, the die runs only 7–8°C above the boiling refrigerant.

And it gets better as chips get hotter. The boiling heat transfer coefficient climbed past 110,000 W/m²K at the highest heat flux tested, with no dryout and no critical-heat-flux wall even at 2.7 kW, more than double a B200. The physics favors this approach at the power levels the industry is heading toward.

That headroom is where the facility water story goes next. Cut 25°C+ of thermal stack down to 8°C and the whole temperature chain relaxes. The paper argues the CDU could run on 60–70°C facility water which is impressive. For reference: an H100 throttles at 87°C as we learned in the previous whitepaper, and the paper calculates you could boil at up to 80°C at the chip. At 60–70°C facility water, chillers are not a contingency anymore but they could be gone from the conversation in any climate on Earth. Also the waste heat stops being a disposal problem and becomes a product: 60–70°C is district-heating grade, and most European countries already mandates waste-heat reuse for new data centers. 60–70°C facility water is not a coincidence on itself because it's the threshold where waste heat stops needing capital equipment to be useful.

So let’s add the results to the table we created in our previous technical series about the Accelsius whitepaper so we can compare them.

Thermal resistance is the most impressive result, so let’s be clear on what it measures. Thermal resistance in K/W is the temperature tax per watt: how much hotter the chip runs than its coolant for every watt it dissipates. At 0.0328 K/W a 1,000 W GPU sits about 33°C above its coolant. At 0.007 K/W the same chip sits about 8°C above it. That’s the whole story in one number.

Each step roughly halves it: 0.0328 → 0.0171 → 0.007. But the two halvings come from entirely different things. The first is a fluid change where boiling absorbs heat at a nearly constant temperature instead of the water warming up as it flows. The second is mechanical by deleting the TIM and the cold plate baseplate.

And here’s the number that explains why that second step is even possible. The paper estimates that at high heat fluxes on small dies, the boiling itself accounts for only about 30% of total thermal resistance. The other 70% is TIM and copper conduction, which is exactly what direct-on-die throws away. Accelsius’ own conclusion: for a chip manufacturer that can etch channels into the silicon, removing those contributions buys “up to 50% warmer facility water.”

Note the condition in that sentence, because it decides everything. The 50% depends on chipmakers changing how they package GPUs.

And something else that the paper mentions quietly is their measured boiling performance which is in line with what other researchers have already published. The advantage here is not proprietary boiling wizardry, it is structural. You get it by deleting the stack, not by boiling better than anyone else. That is what keeps chiller-free operation possible as chip power climbs, and it extends it from cool climates to warm ones.

The ΔT rows are consistent, which is a good sign. Multiply resistance by roughly 1 kW of chip power and you get the temperature rise: 0.0171 × 1,000 W = 17°C, 0.007 × 1,200 W ≈ 8°C. The numbers aren’t cherry-picked from different operating points; they hang together.

We can also compare the results against the Microsoft solution but it should be noted that Microsoft tested on a real GPU while Accelsius tested a copper block, which favours Microsoft on realism. But Accelsius published a specific, checkable thermal resistance while Microsoft published percentages and ratios, which favours Accelsius on transparency.

Our Opinion About The Direct-to-Die Paper

Same treatment we gave the previous whitepaper. This is a lab result on a copper block with cartridge heaters. Not a GPU, not silicon, not a server, not a rack. The headline 0.007 K/W needs fins on the die with either a micro heat sink bonded to the chip or microchannels etched into the silicon. While the results are good, it means the chip manufacturer has to change how it packages GPUs, and that is not Accelsius’s decision. It’s NVIDIA’s and TSMC’s so we don’t expect to see this coming to market in the short term but we like to be surprised. But the good thing is that Accelsius is part of the NVIDIA Inception program and there are pictures circulating on X showing the NVIDIA team working alongside Accelsius, which might include future direct-to-die cooling designs.

PFAS-Free Two-Phase Paper

In April 2026, Accelsius had published a paper, written with three engineers from SK Innovation in Seoul, is called “PFAS-Free Two-Phase Direct-to-Chip Cooling Using a Glide Refrigerant” which I completely missed and I hadn’t seen any comment about this on X either. So let’s fold it in here rather than writing another technical series just about this paper as the technical series is meant to cover more topics than just two-phase liquid cooling.

By now you should know that two-phase cooling only works because the coolant is dielectric, leak it onto a GPU and nothing happens. But dielectric refrigerants are dielectric because they are fluorinated chemicals (also known as PFAS), and Europe is moving toward broad restrictions on PFAS. If the fluid gets regulated, no amount of thermal performance saves the technology.

The paper tests a new refrigerant, R-4101A, newly designated under ASHRAE Standard 34. It’s a blend of R-32, R-152a and R-13I1 at roughly 11%, 30% and 59% by mass. Under the OECD and European Chemicals Agency definition, a PFAS is a fluorinated substance with at least one fully fluorinated carbon carrying no hydrogen, chlorine, bromine or iodine. R-32 and R-152a both have hydrogen on the fluorinated carbon. R-13I1 has a fully fluorinated carbon, but with an iodine atom attached which the definition explicitly excludes. All three fall outside so the blend is PFAS-free.

Then they measured it: same microchannel cold plate, same flow rate, tested against R515B. Once boiling is fully developed, the thermal resistance of the two fluids is essentially identical. You can go PFAS-free without giving up cold plate performance.

What it solves

Two things. The obvious one is regulatory: a shipping-ready path off PFAS-exposed fluids, published before the bear case was even made.

The more interesting one is thermodynamic, and it addresses a weakness we flagged in our review of the whitepaper. In a two-phase CDU, refrigerant condenses at a single constant temperature while facility water warms gradually as it flows. The profiles mismatch, forcing a large approach temperature difference across the condenser and every degree lost there is a degree taken back from what the cold plate that was just earned. This is why 19°C at the cold plate became only 9–14°C at system level. To their credit, Accelsius says so in print: the CDU differential means the cold plate advantage “can shrink.”

R-4101A is their fix. Being zeotropic, it condenses across a glide of about 7.5°C rather than at one temperature. Match that glide to the facility water’s temperature rise and the profiles run parallel instead of pinching, driving the theoretical floor on approach temperature toward zero.

The new risks it adds

The fluid is mildly flammable which is why R-4101A carries an A2L safety classification. Both of Accelsius’ current fluids are A1, non-flammable. A2L refrigerants are now routine in air conditioning, but a data hall is a different conversation involving fire codes, insurers and local permitting authorities. The paper handles this in one sentence about “safety-by-design measures”. This basically means that the operator just should take the appropriate measures to prevent the refigerant to catch fire. For a technology sold on risk reduction, trading a regulatory risk for a flammability risk needs a fuller answer than that.

The PFAS-free status rests on a definitional carve-out, not on the absence of fluorine. Most of the blend is R-13I1, which escapes the definition only because of its iodine atom. Europe’s definition is still being debated, and a broader version could close that door.

Two smaller items: R-4101A needs more power than R515B before boiling is fully developed, so at partial load or on lower-power components it performs worse.

Timeline to product

Nothing here is close to shipping, and the paper doesn’t claim otherwise. It is cold-plate-level work where the condenser benefit is presented as a schematic and a physical principle, with no CDU test and no measured system-level gain. Every benefit is phrased as “potential.”

The realistic sequence is CDU-level validation of the glide-matching gain, then flammability qualification with fire authorities and insurers, then a fluid option on existing hardware. Assuming R-4101A drops into the current IR150 and MR250 without new components, which is unverified. Watch the OCP Global Summit in October, where Accelsius has said new announcements land. A shipping PFAS-free option in 2027 looks plausible but sooner would surprise us, and the A2L question is more likely to set the pace than the thermal engineering.

Market impact: who wins, who worries

Winners

• Accelsius / Innventure ($INV). The obvious one, but note what kind of win: strategic, not financial. This paper repositions Accelsius from “cold plate vendor racing CoolIT” to “the company building the bridge to die-level cooling.” No near-term revenue. Meaningful narrative and IP value.

• Refrigerant suppliers like Honeywell, Chemours, SK and others. R515b and R-401A were engineered specifically for data center cooling. If two-phase scales, at the cold plate or at the die, the refrigerant supplier ships volume either way.

• The two-phase category, including ZutaCore. Every credible demonstration that boiling dielectric handles 200 W/cm² makes the category easier to underwrite. And remember our earlier point that hyperscalers prefer two suppliers before they adopt anything.

• Dry cooler makers and heat-reuse players. At 60–70°C facility water, dry coolers (Carrier, Jaeggi, Güntner etc) replace chiller plants nearly everywhere, and district heating operators get a genuinely useful heat source. The EU regulatory tailwind is real.

• Chipmakers and foundries. NVIDIA gets a path to 3 kW+ TDPs without exotic facility water. TSMC gets a new high-value packaging step. They win either way; the question is which cooling route they bless.

Mixed

• CDU and loop vendors (Vertiv, Motivair, and Accelsius’s own CDU line). Bernstein’s scorecard says direct-to-die does not eliminate CDUs, manifolds or loops. The loop survives; only the last centimeter changes. Whoever’s loop is compatible with the die-level era keeps selling. Water-loop CDUs would need conversion.

• Microsoft / Corintis. Validation that die-level cooling is the endgame but now with a competing route from an unexpected direction. Watch whether the routes converge: dielectric fluid in etched channels would combine the benefits of both.

Losers

• Cold plate merchants, long term. CoolIT’s 15 kW plate is a genuine achievement. But this paper (and Microsoft’s work) says the cold plate as a product category has an expiry date: it gets absorbed into the chip package, and the value migrates to the chipmaker and the fluid and loop suppliers. Boyd, AVC, Auras and CoolIT itself have years of runway and a strategic question waiting at the end of it.

• TIM suppliers. Both die-level routes delete the thermal interface material entirely. A quiet, niche casualty nobody is writing about.

• Chiller OEMs, again. 45°C water made chillers a contingency; 60–70°C water makes them irrelevant for IT cooling in any climate. Johnson Controls, Carrier and Trane already saw their shares decline on one Jensen keynote line in January, this is the same story with another decade of runway. And once more the irony: JCI put $65M into Accelsius. The chiller company bought a hedge on the technology that kills chillers.

Our Call With Accelsius

Our call with the Accelsius team is being scheduled after earnings, possibly as soon as 14 August or the week after. We will report back on the questions we set out in our previous Technical Series piece, and we’ll add a few new ones that came out of these two papers.

The first is simply whether direct-on-die is a product roadmap or a research exercise. If the roadmap is real, it would be interesting to understand who else is at the table, whether NVIDIA, TSMC or a packaging house is involved, who would actually bond the fins or etch the channels, and when they expect this to become part of a GPU design rather than a lab result. We’d also like to know what the optimised jet version delivers, because the bare-die case ran into its ceiling at Blackwell power, and whether that follow-up work gets published too.

On the refrigerant side we’re curious why the die-level study used R515B rather than the cleaner R1233zd(E), and whether those results hold with the latter. And the A2L question is the one we’d most like a straight answer on: what customers have said, and what Accelsius thought is about a mildly flammable fluid in a data hall. Also how the charge limits work out under EN 378 and IEC 60335-2-40, whether fire authorities and insurers have been consulted yet, and whether we can see the safety data sheet. Our sense is that this, rather than the thermal engineering, is what sets the timeline.

Then there’s the competing route. Microsoft, Corintis and TSMC are getting to the die with sealed channels and ordinary water. We’d like to hear why a chipmaker would choose open dielectric impingement instead because that decision is not Accelsius’ to make, and it may be the single biggest variable in this whole story.

Takeaway

Two weeks ago we wrote that the Accelsius whitepaper was a strong opening argument in a trial that hasn’t happened yet. These two papers are a different kind of document. Neither is a benchmark against a competitor. They are flags planted on where the industry could end up and between them they quietly answer the two questions hanging over every two-phase investment case. What happens when cold plates themselves become obsolete? And what happens if Europe regulates the refrigerant?

The short answers: Accelsius intends to own the physics at the die, and they had a PFAS-free fluid on the bench months before anyone published the objection.

That is the bull case but both results are lab work. The direct-on-die numbers need a chip manufacturer to change how GPUs are packaged, and that is NVIDIA’s and TSMC’s decision, not Accelsius’. The PFAS-free fluid is mildly flammable, and fire codes and insurers move slower than thermal engineering. Microsoft and TSMC are pursuing a competing route to the die that lets water in through sealed channels. None of that is fatal. All of it is unresolved.

What we do take from it is this is that Accelsius publishes. Three papers in three months, in IEEE and ASME venues, with named co-authors, methodology you can argue with, and their own system-level weakness written down in print. We have also spent this week reading a competitor’s material that does not contain a single thermal resistance figure. That contrast tells you which company is engineering-led.

The tweet that started all this didn’t get much traction, probably because most people don’t feel like reading engineering papers, but luckily we do.

Disclosure

We hold, or have held, a position in Innventure (NASDAQ: INV). This article reflects our opinion, is for informational purposes only, and is not investment advice. Do your own research. Sources: Accelsius, “Direct On Die Two Phase Cooling Approach for High Power GPUs” (Narayanan, Wang, Ozguc, Whitaker, Moore, Bonner — June 2026); Accelsius “Warm-Water Ready” whitepaper (July 2026); Bernstein, “Liquid Cooling in Data Centers” (16 July 2026), brief excerpts under fair use; Microsoft/Corintis in-chip microfluidics announcements (Sept 2025); Honeywell Solstice N15 (R-515B) documentation.

1) KEY METRICS / DATA POINTS

Let’s start with the jobs report on Friday. It was the print of the week and it slightly rearranged the conversation around the Fed. Nonfarm payrolls fell by 23,000 in July, well below the 83,000 consensus. That was also below every forecast on Bloomberg. Unemployment ticked down to 4.1% from 4.2%, but only because 264,000 people dropped out of the civilian labour force. Participation fell to 61.4%, the lowest in over five years. Wage growth softened to 3.2% annually, the weakest since May 2021. And here is the number that matters most: May and June payrolls were revised down by a combined 103,000. The three-month picture is meaningfully worse than the earlier headlines told us. Local government education shed 50,000 jobs. Retail lost 19,000. Health care kept adding. Private-sector payrolls rose 30,000, offsetting a 53,000 government decline.

This labour market has stopped softening at the edges. The three-month average of payroll gains has fallen to 35,000, which is below where most economists put the break-even rate. The reason unemployment did not spike is that the labour force is shrinking almost as fast as employment. Immigration curbs and demographic drift are doing most of that work.

Now Japan, which was the more surprising development of the week. On Monday August 3, the US Treasury and the Bank of Japan carried out their first coordinated yen-buying intervention since 2022. The trigger was renewed yen weakness against the dollar. The deeper worry was Japanese government bond yields spilling straight into US Treasury markets. The 10-year JGB yield touched 2.82% during the week before pulling back on the Iran news. The Ministry of Finance confirmed access to the Fed’s FIMA repo facility at the same time, which lets Japan raise dollar liquidity without dumping Treasuries. The message is that Washington no longer treats Japan’s currency and bond markets as a purely Japanese problem. They are firmly a US concern too. A persistently weak yen risks forcing Japanese institutional investors to unwind their roughly $5 trillion in foreign assets. That unwind would land primarily in US Treasuries at exactly the moment the long end is already under pressure.

Q2 earnings passed their peak this week. S&P 500 profit growth is running near 37.9% year-on-year, with 61% of the index reported and 64% of those beating consensus by more than one standard deviation. Strip out one-time gains from big tech stakes and EPS still grew 26%. That is the strongest since 2021. AI infrastructure companies accounted for roughly a third of that gain. The hyperscalers were the story. Alphabet raised its 2026 capex guidance to $195-205 billion. Combined hyperscaler capex is now projected at $760 billion this year, up from $750 billion just weeks ago. Amazon delivered a clearer capex-payback framework than expected and the market rewarded it. Microsoft delivered strong Azure growth. Meta was punished for what looked like capex discipline, which tells you something about the current mood. Alphabet’s Q2 capex alone came in at $44.9 billion, more than double a year earlier. Free cash flow swung to negative $5.9 billion even as operating income hit $40.8 billion.

The RIA Advisors analysis making the rounds this week gave the hyperscaler dynamic a number worth sitting with. In 2026, the five biggest hyperscalers will spend roughly $760 billion on AI infrastructure. They will expense only $211 billion. The other $549 billion is capitalised and sits on future income statements as depreciation. Combined net income across the five is projected to rise 25% this year to about $506 billion. Combined free cash flow is projected to fall 91% to roughly $16 billion. The market is being asked to trust that the revenue this capex generates will arrive fast enough to absorb the depreciation drag when it lands.

On Iran, the pattern from the summer held. Trump’s weekend post cancelling the strike gave way to renewed Omani-mediated negotiations. Oil settled in the $80 to $85 range. WTI was around $81 by Friday. Brent near $85. There were no major strikes in either direction this week. That is the closest thing to peace we have seen since the July 8 collapse. Whether it lasts is the question.

The Fed is still 3.50% to 3.75%, so it did not move. What changed was the market pricing around September. Before Friday’s jobs report, markets had roughly 42% odds of a hold in September, with the rest split between hikes and other actions. By Friday afternoon, cut odds had risen sharply. Several bank strategists moved cuts to their base case. The tension between an inflation problem and a labour market that has clearly stopped growing is now sharp enough that the September decision looks meaningfully different than it did a week ago.

2) AAIG’S PERSPECTIVE

We want to walk through what all this means together. These threads connect in ways that get lost if you look at each release on its own.

The jobs report is what changes the near-term Fed calculus. For most of 2026, the Fed has been fighting an energy-shock inflation problem while the labour market held up well enough to justify caution on cuts. That story is no longer available with the recent releases. A negative payrolls print, a three-month average of 35,000, and downward revisions of 103,000 clearly shows the cracks in the labour market. The dual mandate is now visibly pulling in two directions. Inflation is still elevated and Tuesday’s CPI will likely show it moving higher on the reversal in energy. Unemployment is being held down artificially by a shrinking labour force. Immigration policy is now the swing variable. If the labour force stops shrinking, the unemployment rate will rise mechanically, and it will look worse in a hurry.

Markets have responded by pricing in more cuts. We would be careful about drawing a straight line from Friday’s weak print to a September cut. There is a distance between what markets want and what Warsh can actually deliver. He has three hawkish dissenters on his committee who voted for a hike two weeks ago. Cutting in September, against that internal opposition, with headline inflation likely printing above 3.5% on Tuesday, would be a big move for a chair still in his first six months. The more realistic path is that Warsh signals openness to cuts if the data continues in this direction, holds September, and starts positioning for October or December. Market pricing may be running ahead of what the Fed can actually do without cracking internal consensus.

The Japan intervention is a slower burn but potentially the more consequential development over the medium term. The US does not run joint currency interventions casually. This happens when Washington believes a foreign country’s monetary and fiscal choices are directly threatening US financial stability. What is different this time: the US worry is not really about the yen. It is that a weak yen and rising JGB yields together could force Japanese institutions to sell US Treasuries at scale. Japan owns roughly $1.1 trillion in Treasuries. If yen weakness accelerates and JGB yields keep climbing, some of that could get forced back home. The FIMA repo access is designed specifically to prevent that outcome. That tells you exactly what Washington is worried about.

For US investors, the practical concern is that the long end of the Treasury curve is now vulnerable to developments that are not fully within US control. If the BOJ hikes further, or if Prime Minister Takaichi pushes ahead with an aggressive fiscal package this autumn, US 30-year yields can move higher regardless of what the Fed does. That is a real risk to duration positioning. It has been building all year without getting proportionate attention.

The Q2 earnings picture is more layered than the headline numbers make it look. 37.9% profit growth is a strong result. Cloud revenue at the hyperscalers is real evidence that AI capex is beginning to translate into revenue. The free cash flow dynamic is worth sitting with though. Capex at this scale, spread over multi-year depreciation schedules, makes current earnings look stronger than the cash reality. The gap between $760 billion of AI capex being deployed in 2026 and $211 billion being expensed this year is real. Eventually it shows up either as compressed earnings, if depreciation lands before revenue catches up, or as sustained earnings, if the revenue does arrive in time. Which one materialises will determine a lot about equity returns over the next two to three years. Nobody knows the answer with confidence right now.

What we do know: this is not an argument about whether AI is real. AI is real. Cloud revenue is real. Productivity gains are showing up in specific sectors already. The question is whether the pace of capex deployment is calibrated to the pace of revenue realisation. Hyperscaler management teams sound confident. Investors are starting to ask whether that confidence is calibrated correctly, or whether it is a habit built up over a period where nobody got penalised for spending too much on AI. Alphabet’s post-earnings weakness on the $200 billion capex guidance is the first sign the market is beginning to price this discipline question.

The link between all three (earnings, Japan, jobs) runs through the same variable: term premium in the long end of the curve. Weak jobs strengthen the case for Fed cuts and would normally pull yields lower. Hyperscaler capex financed through equity and debt issuance adds Treasury supply pressure. Japan intervention exposes the risk that Japanese buyers may not be there for that supply. The AI boom, the labour softening, and the Japan situation are all being resolved through the same channel. None of the resolutions is clean.

On Iran, the market’s willingness to hold oil in the $80 to $85 range this week is the closest thing to a vote of confidence in the diplomatic process we have seen since June 17. It is a conditional vote. It depends on the Omani channel producing something concrete in August, either an IAEA access agreement or a formal reopening protocol for the strait. Without one of those, the pattern that has held all summer, talks, breakdown, oil spike, private climbdown, talks resume, will repeat. The base case is that the cycle continues into September and possibly through Jackson Hole. A real resolution requires one of the two sides to accept a compromise neither has been willing to offer so far.

3) WHAT TO WATCH

Three things over the next two weeks will shape everything else.

The first is the July CPI release on Tuesday August 12. This is the last major inflation print before the September FOMC meeting. Gasoline reversed sharply in July, so headline CPI will move higher than the 3.5% June print. Somewhere in the 3.8% to 4.0% range is the reasonable range. Core CPI is the number that matters. If it prints at 0.2% or 0.3% monthly, the case for holding in September stays intact even with the weak jobs report. If core comes in at 0.1% or flat, the case for a September cut gets much harder to resist. We expect core to firm up modestly rather than moderate further, based on the shelter and services trends. In that scenario, the Fed is genuinely caught between the two sides of its mandate.

The second is Jackson Hole on August 27-29. This is Warsh’s first appearance at the annual symposium as chair. He has been consistent about not providing forward guidance. The labour market data may force his hand anyway. If he uses Jackson Hole to explicitly acknowledge that the labour market has weakened enough to warrant policy action, markets will price a September cut as near-certain. If he holds his line and reiterates data dependence, September stays open and the CPI print becomes the pivot.

The third is whether Japan’s intervention holds. The yen strengthened sharply on Monday and has stayed off its lows through the week. JGB yields pulled back on the Iran news. The underlying pressure has not been resolved. Takaichi is expected to push a fiscal expansion package in the autumn. If the BOJ does not hike rates in response, the intervention will need to be repeated. A second coordinated intervention within the next month would be a signal that the situation is deteriorating faster than policymakers can manage. If Japan and the US can hold the yen at current levels through August without further action, some of the Treasury market pressure will ease.

The thread across all of this is that the interconnections between markets, geographies, and policy tools are tighter than they have been in years. A jobs miss in the US moves Fed pricing. Fed pricing moves Treasury yields. Treasury yields move the yen. The yen moves Japanese institutional flows. Those flows move Treasuries again. Every one of these variables is in motion at the same time. Each has a real capacity to surprise. Hyperscaler capex, oil, and Iran sit on top. The market is not currently pricing any of them individually as dangerous. The risk we care about is not any single one going wrong. It is two or three going wrong together, and the interactions becoming the story.

That does not mean disaster is coming. It means the range of plausible outcomes is wider than at any point since March, and the tails are fatter than the current implied volatility in equities suggests. Position accordingly.

Small reminder: July CPI release on August 12!

Thanks for reading this 17th pulse, and see you at the next one!

AAIG is a research collective built on deep work, shared scrutiny, and quality over quantity. We treat research as a product, not content to fill a feed. By combining the expertise of multiple specialists, we find true edge where others just see noise. No hype, no daily trades. Just high-conviction ideas.

This is Asset Alliance Investment Group: Learn more on substack.com/@aaig

Disclaimer: The research and analysis published by Asset Alliance Investment Group is for informational purposes only and does not constitute financial, investment, or professional advice of any kind. All investment decisions carry risk, and recipients are strongly encouraged to seek independent, qualified financial advice before acting on any information contained herein.

I have moved the Rubin Ultra de-spec to the top of my memory notes, and not because it is bearish. It is a important read on the latent HBM demand thesis. So I want to walk through what it actually tells us, and then I want to walk through every reason it might be wrong, because I have spent the last weeks stress testing this against my own thesis rather than looking for reasons to feel good about it.

Start with what is not in dispute as Jukan pointed out correctly. Nvidia’s incentive is to ship as many Rubin racks as possible, and lowering HBM content in favour of SOCAMM does exactly that, because of the wafer usage.

Every stack of HBM consumes wafers that could have gone to the module memory sitting on the other side of the rack, and a rack that is short of module memory does not ship at all. From the supply side, the memory players sell all their bits anyway, just spread across more units, and they pick up some incremental module-memory revenue on top. There is no margin conflict here. Both sides win. Nvidia ships more racks, and the memory makers sell everything they can make.

But winning is not the same as getting what you wanted. Nvidia is optimising under a constraint it did not choose. It is making the best of the situation. And that distinction is the whole article.

What the buyer told us without meaning to

Because in this case Nvidia has effectively handed us both numbers.

What it originally wanted is on the record: the announced 1 TB Rubin Ultra configuration. What it is reportedly settling for is around 192 GB, still in testing and not finalised. That is roughly a fivefold reduction, and it is the difference between design intent and delivered reality, published by the buyer itself in two separate acts.

That is revealed behaviour, not a survey, and it is the basis for how I estimate wish-list demand against the small deficit the sell side currently models. Most latent demand is inferred. Original intent was announced and then reduced.

In June, Jensen visited SK hynix at Computex, signed their first HBM4E wafer sample, and left a message on it asking them to make more. clear signal from Jensen.

Now let me be precise about the mechanism, because the usual description of this is wrong and the difference matters.

A part of the people say HBM simply got too expensive for the current architecture. In a shortage, price is not the thing that binds. Availability is. Nvidia is not spec. down because the bill got large. It is specking down because the cost that moved is not the price of HBM. It is the opportunity cost of the wafers standing behind it. And an opportunity cost does not disappear if prices come down. Simple wafer math.

Why the want does not decay in my eyes.

That underlying want does not decay while it is being suppressed, and it does not decay because of NPO either. The workload driver keeps rising. Context windows have grown by orders of magnitude and are still growing. Whatever wanted 1 TB in 2027 will want more than that later. So even if supply reached parity, a part of unserved demand accumulates through the whole de-spec period and releases at the next platform. a potential memory increase as a drop-in when it fits the existing mechanical, thermal and power budget, and forces a platform refresh is also unlikely. And doesnt impact the thesis. Going from 192GB to a 1TB-class configuration with 16-high stacks at relaxed package height is very unlikely to fit.

So what happens at the next architecture? Here I think most people have the mechanism backwards.

The usual framing is that the compute upgrade pays for the extra memory. I do not read it that way. Token generation is bandwidth bound, not compute bound. Add compute without the memory to feed it and you have bought an accelerator that waits. There is a published measurement of exactly this: push the working set one tier down the hierarchy and nearly all the latency ends up sitting in transfers, while the accelerator runs at a small fraction of its power envelope. That is a machine paying for compute it cannot use.

So the next generation does not carry more HBM because it can suddenly afford to. It carries more because the compute it is being sold on is not fully usable without it. Memory is not competing with compute for the budget. It is the thing that unlocks the budget, and of course you pay for the stage that sets your output.

The generational data says the same thing from the other side. Compute per accelerator has been growing two to three times per generation, while bandwidth per memory stack grows maybe half again. That gap is the memory wall, and it widens every cycle. Just holding the bandwidth-to-compute ratio constant requires more stacks per accelerator each generation, not fewer. Which means the real ceiling on HBM content is not demand at all. It is geometry, how many stacks physically fit around a package. And a binding geometry cap is an argument for more accelerators, not for less memory.

Can the architectures just walk away from HBM?

This is the question I spent the most time on, because it is the one that would actually end the thesis. I stress tested it against every alternative on the table, and the physics keeps answering the same way.

Start with SRAM, because it is the cleanest case. The dream solution has always been to put the memory on the logic die itself. The problem is that SRAM has effectively stopped scaling. Logic density still improves meaningfully every node. SRAM bit cells have barely moved across the last two nodes, a few percent against roughly forty for logic. Which means every new node makes on-die memory relatively more expensive, not less. The escape route that would most cleanly remove external HBM looks like it is closing. and it is closing for physics reasons.

Then pooling and tiering, and here is the point almost everyone misses. HBM is bought in stacks, and every stack delivers bandwidth and capacity together, in fixed proportion. Nearly every substitution story you will read is a capacity story: pool the cold weights, tier the cache, offload to flash. But if bandwidth is what binds, and for token generation it is, then removing capacity requirement removes zero stacks. The buyer still needs the stacks for their bandwidth and simply ends up with spare capacity. And this is important i will come back to that later. When you run the interconnect numbers. Pooled memory over the current standard moves on the order of a hundred gigabytes per second per link. One HBM stack moves on the order of two terabytes per second. That is more than an order of magnitude, and it is a property of the interface, not a maturity problem that another software release fixes. Pooling is real, and I watch its deployment dates. What it serves is the capacity-bound edge of the market. It does not reach the bandwidth-bound core.

The LPDDR designs carrying no HBM at all are real too. Shipping, with named customers. I model them as a deduction rather than pretending they do not exist. But look at what they serve: capacity-bound inference at lower throughput. They grow the accelerator market at its edge. They do not displace the centre.

And I want to be honest about the shape of what I just gave you. That is not four arguments. It is one argument with four supports: everything cheaper than HBM is an order of magnitude short on bandwidth, and the gap is physical.

But what if

There is a version of this thesis that says every buyer wants more HBM than it can get, and that version does not survive contact with the record.

At least one major custom-silicon programme has specced its current generation on the previous HBM generation across both of its configurations, and the read there is a deliberate yield and cost choice rather than a supply constraint. That buyer did not want more capable memory. It wanted cheaper memory. In this case Google TPU.

But look at why, because the reason is structural rather than a difference in taste. An internal programme is optimising a task-specific workload. It knows exactly what it will run, it controls both the model and the silicon, and it can spec precisely to that. So tailor made to their own workflow.

Nvidia is selling into a workload it has never seen. Frontier training, frontier inference, and sovereign build-outs like the Japan programme, where the buyer itself does not yet know what it will be running in three years. You cannot spec down for a workload you have not seen. You want your best product there.

What would change my mind… and then Jevon kicks in

I would rather write this down now than be asked for it later.

Everything I said about physics defends against hardware substitutes. It defends against nothing on the software side. The fastest-growing piece of memory demand is the KV cache, and its size is a direct function of the attention mechanism the models use. Today’s frontier models use full attention, which makes that cache grow with context length, and that growth is a big part of my demand case. If the frontier moves to sub-quadratic attention and holds quality on long-range work, the cache stops growing the way I assume. No new silicon required, and one model generation to propagate across the installed base. It has not happened. The hybrids that exist still show quality gaps at frontier scale, and history says a cheaper cache gets spent immediately on longer context. This is Jevons paradox. And all major names strengthen this direction in commentary. We will have much bigger models in the future.

The second is the design question. If the next architecture also lands below its own design intent, under looser supply, then the reduction was a preference and not a supply response, and my reading of this whole episode is wrong.

The third is efficiency at the fleet level. If the next generation makes each accelerator productive enough that buyers need fewer of them for the same tokens, then content per unit rises while unit counts fall and total bits go nowhere. Same answer as above. Every efficiency gain in this field so far has been reinvested in bigger models and longer context. And will be for the foreseable future. Which equals more HBM demand

None of those “tripwires” is shown yet. Until one does, what the buyers wanted is a higher HBM spec than what they were able to get.

And that is what makes the case for the memory suppliers, on the shortage, the pricing power and the earnings durability, a much longer-term one than the next 4 quarters potential earnings visibility. Their wishlist has not changed.

Final conclusion

Almost everyone modelling this memory cycle does the same thing. Supply on one side, order demand on the other, subtract, publish the deficit. Which of course gives you a pretty small shortage.

But order demand in a supply-constrained market is a censored number. It is simply what buyers ask for knowing what they can get. So it tells you a lot about how the shortage is being administered, and nothing about what happens when supply meets the wish list.

The wish list is the only thing that speaks to the years nobody can really see. And building it means you have to model and understand why the memory is needed at all. That is what I explained above. The bandwidth per token. The cache that keeps growing with context. And with that, the wall that widens every generation. Not what was allocated, but what would be bought.

And if you look at the wish list, your question for seeing beyond 2028 should be simple. Either the new capacity meets a satisfied market, or an unsatisfied one.

With all the points I have presented, and the behaviour of Nvidia, the choice to optimise for bandwidth so you can get the maximum out of your newer generation, which you then have to drop because you need more SOCAMM, for me that signs unsatisfied.

Thank you all for reading.

Markos

I am genuinely pleased to see a well-known brokerage finally publish an analysis that aligns closely with the thesis behind my investment in LPKF.

In a way, it feels like I recognized the value of glass substrates too early and paid the price for it. In any case, today’s topic is CITIC Securities’ report and how it compares with LPKF.

Below, I go point by point through where the two views converge and where they don’t: the commercialization timeline, the 2030 penetration rate, and which process steps CITIC flags as the critical bottlenecks. All of which map onto the exact processes LPKF is targeting.

I also address a distinction that’s easy to get wrong: CITIC’s TAM and LPKF’s TAM are not measuring the same thing, and conflating the two will lead you to the wrong conclusion about what this report actually implies for LPKF’s revenue opportunity.

Finally, I lay out the three risks I think investors should keep in mind before treating CITIC’s numbers as validation of LPKF’s upper-case scenario, along with the specific indicators I’ll be watching over the next few quarters to see whether this thesis is actually playing out.

Read more

In this post I will break down how you need to compare these notes and what sell-side doesnt show.

First, These past days if seen numerous sell-side note on memory in general. And these summaries of sell-side research increasingly quote HBM in exabytes as if everybody is measuring the same thing…

But the thing is they are not.

Depending on the report, “HBM volume” can mean:
• Accelerator demand multiplied by the HBM currently specified per GPU
• Supplier bit shipments
• HBM consumed months later in completed systems
• Actual wafer input or wafer-equivalent production capacity
• Qualified shippable output after stacking, testing and customer qualification
• Revenue TAM, calculated as shipped bits × ASP

These can all be valid metrics don’t get me wrong. But when you put them on one chart without reconciling the definitions properly as a reader and you can create an apparent shortage, surplus or growth rate that does not physically exist.

Imo the clearest example is 2027.

UBS’ latest publicly summarized demand forecast is 58.7bn Gb, or 7.338EB. Its latest publicly summarized supply volume is 59.087bn Gb, or 7.386EB.
Within UBS’ own served-market model, aggregate HBM therefore looks almost perfectly balanced.

Daishin, however, forecasts 51.0bn Gb of 2027 annual shipments, or 6.375EB.
Subtract Daishin’s shipments from UBS’ demand and the apparent shortage is 0.963EB, or 13.1%.

But that is not a clean shortage calculation. It combines one research house’s demand forecast with another house’s shipment forecast, using different capacity, productivity and timing assumptions. And be very aware of that when you use it as a reference for yourself.

The 13.1% is therefore not proof that the physical shortage is exactly 13.1%. It is an example of how an apparently precise shortage can emerge from two researches that were never designed to be directly compared. Even tho we like to compare them.

Earlier UBS research makes the timing problem even clearer.

It separated 53.9bn Gb of 2027 end-consumption (6.738 EB) from 58.3bn Gb of supplier bit shipments (7.288 EB) and cited an approximately 3.5-month lag between supplier shipment and consumption inside a completed rack.

So 7.288 EB minus 6.738 EB is not automatically a 0.55EB surplus. Much of that difference is HBM moving through the production and inventory pipeline.

The same definitional problem appears on the wafer side.

For end-2026:
• TrendForce says HBM represents approximately 22% of top-three DRAM wafer input
• UBS has been publicly summarized at 500k equivalent 12-inch WPM, or 25% of industry capacity
• Daishin reports 535k WPM of HBM-designated production capacity, representing 31% of its 1.73m WPM total

All three figures are presented as indications of HBM’s wafer footprint.
But again they are not the same measurement.

TrendForce describes actual or forecast wafer input. UBS describes normalized equivalent capacity. Daishin describes product-family production capacity. They use different denominators, timing points and normalizations.

Even the calculation “end-WPM × 12” is misleading during a capacity ramp. Obviously December’s capacity did not exist in January.

Daishin also models a 5–6 month HBM production lead time. That means 2027 shipments depend heavily on wafer capacity operating during the second half of 2026 and first half of 2027—not simply on the capacity available at the end of 2027.

Once the definitions are separated, you land on the second issue: the underlying forecasting assumptions.

Daishin publishes quarterly capacity, annual shipments and the reported production lag. Shifting its wafer capacity by that lag implies almost identical shipped output per capacity-wafer in 2026 and 2027: approximately 7.6–7.7Tb.

Applied to the relevant 2027 wafer base, that points to roughly 6.3–6.4EB of supply.

UBS reports 7.386EB.

To reach UBS’ number from the same wafer base, output per wafer would need to be approximately 17% higher. Alternatively, the average wafer capacity contributing to 2027 shipments would need to be approximately 17% larger.

That does not necessarily mean UBS is wrong. UBS may be assuming better yields, higher utilization, denser dies, a different product mix, more wafer allocation or a shorter production lag.

That could be completely fine in terms of modelling btw.

This is not a definitional discrepancy. It is an underlying assumption gap that becomes visible only when you really have seperated the definitions

But There is also a deeper issue: reported “demand” is often already supply-shaped and underwrites the actual demand of memory.

The original Rubin Ultra concept carried 1,024GB of HBM and applied to Ultra, not standard Vera Rubin. UBS later modeled 768GB. A reported dual-die redesign implied approximately 384GB. The latest public summary of SemiAnalysis reporting, still unconfirmed by NVIDIA, points to a 192GB mainstream variant.

The complete 1,024→192GB comparison is not apples-to-apples because the underlying architecture changed. The cleaner reported comparison is the dual-die configuration moving from approximately 384GB to 192GB.

For every one million accelerator packages, that reduction removes 0.192EB from specified demand before anyone calculates the shortage.

That does not automatically mean total HBM demand falls by 0.192EB. A cheaper or easier-to-produce accelerator could sell in higher volumes and offset part of the reduction.

But it does mean that a sell-side demand model using the new specification no longer shows the amount of HBM previously planned per package. And this is where I have a different opinion and addition on that. When you listen to the leading companies and talk with people in the industry every player wants more HBM per architecture.

If memory availability, yield, packaging, power or deployment constraints contributed to that reduction, part of the original demand disappeared from the visible market before the shortage was measured. You follow me?

This is why I call it the HBM wish list.
Sell-side “TAM” generally measures the served market under the latest shippable product specifications. It does not necessarily measure unconstrained technical demand before product spec-downs.

Daishin’s own TAM demonstrates why this distinction matters.
Its modeled HBM revenue rises from $55.0bn in 2026 to $109.8bn in 2027 an increase of almost 100%.

But physical shipments rise from 36.3bn to 51.0bn Gb, or from 4.538EB to 6.375EB: growth of approximately 40.5%. Then revenue nearly doubles because modeled bit volume grows by approximately 40% while blended ASP increases by approximately 42%. Revenue TAM doubling does not mean physical HBM demand doubled.

The point is not that UBS, Daishin or TrendForce must be wrong. They are often answering different questions.

If you want to measure that wishlist (for a big part you can call it latend demand because it compounds because wish is not fulfilled) a correct framework keeps five separate watchpoints:
• Latent or unconstrained wish-list demand
• Specified accelerator demand after product decisions
• Supplier bit shipments
• Qualified shippable supply from the wafer ramp
• End-consumption after packaging and pipeline delays

Before citing an HBM shortage, every number should carry its source, forecast vintage, unit, timing point, supply-chain stage and underlying product specification.

Recent DigiTimes reporting, as summarized by Jukan, illustrates exactly why this distinction matters.
The report says 2027 DRAM and HBM capacity is already fully booked, while manufacturers typically deliver only 60–70% of the volumes buyers originally requested.

In other words, the volume that ultimately gets allocated and shipped can be materially lower than the volume customers initially tried to secure.

That does not give us a clean HBM-only exabyte number. It also does not mean every requested unit represents uncontaminated end-demand: scarcity can encourage precautionary ordering, duplicated requests and attempts to defend future allocation.

But structurally, it validates the HBM wish-list framework.

Requested volume, booked capacity, allocated supply, supplier shipments and final consumption are not the same thing.

Most sell-side TAM and shipment models show the market that suppliers can actually serve. They do not fully show the larger amount customers initially wanted, nor the demand removed through allocation, product redesigns and specification reductions.

The visible HBM market is therefore not necessarily the full market customers asked for. It is the market that available supply can serve. Big gap. SK hynix CEO pointed out “5 times” more demand then they can deliver. Thats the wishlist and it impacts the durabilty and timeline of the memory thesis. I estimate those in the upcoming HBM research.

Thanks for reading!

First of all, welcome back everyone! Three weeks have passed since our last pulse due to my well-deserved holiday. It has been three weeks worth writing about though, as the story that had felt like it was settling in mid-June has come apart. The ceasefire collapsed on July 8 when Trump declared it “over” at the NATO summit, and by the end of July oil had posted its strongest monthly gain since May. Sandwiched in between, the June CPI release delivered the softest inflation print in months, arriving at exactly the moment when the conditions that produced it were already reversing. The Fed held rates at its July 29 meeting, but three officials dissented in favour of a hike, making it the most divided FOMC decision Warsh has faced as chair. And late last night, in true Trump fashion, the President took to Truth Social to announce he had called off a planned large-scale strike on Iran, saying the “perimeters of a deal” had been agreed and that Iran and other Middle Eastern countries had asked him to hold off. Each of these events would be significant on its own. Taken together, they mark a real shift in where this cycle is heading. Let’s walk through what happened, what it means, and where we think this is going.!

1) KEY METRICS / DATA POINTS

We should start with the ceasefire, because it collapsed before any of the data came out, and it colours how we read everything else in the Pulse. Trump used the NATO summit stage on July 8 to declare the framework “over” and to describe negotiating with Iran as “a waste of time.” That announcement followed three days of renewed strikes on both sides. Iran hit commercial ships in the strait on July 6 and 7, the US retaliated with strikes on Iranian military targets the following days, and the President raised the possibility of reimposing the naval blockade. Oil moved immediately, with WTI jumping 4.4% to $73.52 and Brent 5.2% to $78.02 on July 8. By July 29, after Iran launched ballistic missiles at US forces in the region, Brent gained another 7.9% in a single session to close at $90.74 while WTI settled at $84.46. By the end of the month, Brent had closed above $88, capping the strongest month for oil since May.

The pattern within that broader move is worth looking at. On July 27, Iran privately told the Trump administration through mediators that it “made a mistake” attacking commercial ships and wanted to return to talks. Oil dropped 8.7% on Brent and 7.5% on WTI in a single session as the market priced a return to negotiations. Two days later, Iran launched missiles at US forces and oil ripped straight back up. This is the cycle we described in our last pulse, but compressed to the point where escalation, private climbdown, negotiation, and re-escalation now happen inside the same week.

And then this weekend brought another turn. On Saturday August 1, Trump posted on Truth Social that the US military was “locked and loaded and ready to go against the Islamic Republic of Iran, at levels of Military Terror, Strength, and Power not seen since World War II,” before adding that he had agreed to cancel the planned strike because Iran and other Middle Eastern countries had asked him to hold off, and because “the perimeters of a deal has been agreed to.” He said the emerging framework would include the “Immediate, Complete, and Total OPENING OF THE HORMUZ STRAIT” and the end of Iran’s nuclear threat. Israel, according to Trump, agreed to hold off as well. This is significant not because we should take the framing at face value (the outline is vague, the deliverables are exactly what has been in every deal iteration since April), but because it tells us the same pattern is repeating: the President moves to the brink, Iran and regional partners request a pause, negotiations resume. What is different this time is the scale of what Trump said was cancelled, a strike on Iran’s energy infrastructure that had reportedly been prepared jointly with Israel. That is not a small thing to walk back.

Underneath the geopolitics, the CPI release on July 14 delivered the most striking data point of the three weeks. Headline CPI fell 0.4% on the month, the largest single-month decline since April 2020, bringing the annual rate down from 4.2% in May to 3.5% in June. Core CPI was flat on the month, with the annual rate down to 2.6% from 2.9% in May. Energy fell 5.7% on the month, its biggest monthly drop since April 2020, driven by a 10% decline in gasoline prices during the brief window in mid-June when the strait was functioning. The framing that mattered came later that day, when Warsh delivered congressional testimony and told the committee: “There might be some that look at this morning’s data and say, ‘Oh, mission accomplished, everything is swell,’ that is not my view.” That single sentence tells you more about where the Fed is going than the print itself, and we will come back to why.

The June jobs report was released ahead of the ceasefire collapse on July 2, and it was soft. Nonfarm payrolls rose just 57,000, well below the 110,000 consensus and the lowest gain in four months. The unemployment rate ticked down to 4.2% from 4.3%, but that improvement came from a shrinking labour force (down 1.1 million since January) rather than from actual job creation. Leisure and hospitality shed 61,000 jobs. Health care, social assistance, and professional services carried most of the gains. Real average hourly earnings rose 0.8% in June, but that was flattered by the negative CPI print. April and May payrolls were revised down by a combined 74,000, meaning the underlying trend has been softer than the earlier headlines suggested.

The July 29 Fed meeting was Warsh’s second as chair and, by his own description, produced “a good family fight.” The committee voted 9-3 to hold at 3.50% to 3.75%, with the three dissenters (Beth Hammack, Neel Kashkari, and Lorie Logan) all voting for a 25 basis point hike. Governor Waller voiced concerns about inflation but voted with the majority. The statement was again short and stripped of forward guidance. Markets sold off during Warsh’s press conference: the Dow fell more than 840 points, the S&P 500 lost 0.6%, and the 30-year Treasury yield rose more than 9 basis points to 5.19%. The CME FedWatch tool now puts the probability of a straight hold in September at just 41.9%, up from 24% the day before, meaning markets are pricing better than even odds of some action other than a hold at that meeting.

2) AAIG’S PERSPECTIVE

The three weeks split cleanly into two stories moving in opposite directions, and understanding both is the key to positioning correctly from here.

The first story is that June’s inflation data was very good. Core CPI at 0.0% monthly and 2.6% annually, combined with the sharp decline in headline, is exactly what the Fed wants to see, and in a different world would be the release that opens the door to rate cuts. Warsh’s testimony on July 14 was the first signal that he does not view this world that way, and the July 29 meeting confirmed it. The reason is that the June data was written in a window that no longer exists. Gasoline fell 10% in June because the strait was functioning for two weeks between the June 17 MOU and the July 8 collapse. That window has closed. Oil finished July more than 20% above where it started, July gasoline prices will reflect that, and the July CPI print due on August 12 will show it. The June data flatters a picture that the market already knows is not the current picture, which is why the CPI release produced almost no lasting rally in either equities or bonds despite being so soft on the surface.

The second story is that the Fed is now genuinely and visibly split, and the direction of the split is hawkish. Three dissenting votes for a rate hike at a single meeting is unusual. What is more unusual is that Warsh, whose personal views are broadly dovish and who has been consistent about wanting room to ease, allowed the meeting to end without providing markets with any clarifying forward guidance. The statement removed easing bias language. The press conference confirmed there is no soft inflation target. Warsh explicitly noted that tightening financial conditions in markets are doing some of the Fed’s work for it, which is code for saying rates may need to stay here or go higher, and the equity and bond markets sorting that out on their own is not something the Fed will interfere with.

The market read this correctly. The sharp move in the 30-year yield tells you where the pressure is. The long end is pricing higher-for-longer with a rising probability of a hike, and the FedWatch tool’s move from near zero six weeks ago to nearly 60% odds of some action beyond a hold at the September meeting is a genuine repricing rather than noise. The trajectory matters more than the level, and the trajectory is clear.

The jobs data adds a layer that complicates the picture. A 57,000 print in June is not consistent with a labour market that needs the Fed to be aggressive, and the shrinking labour force is masking softness that would otherwise show up in the unemployment rate. The three-month average of payroll gains has fallen to 92,000, which is below what most estimates put as the break-even rate needed to keep unemployment stable given labour force dynamics. If the labour force stops shrinking (immigration curbs are the main driver), the unemployment rate will rise mechanically, and that is not a scenario the market is currently pricing.

This creates a genuine tension for Warsh going into September. He has three dissenters on the hawkish side of the committee, a softening labour market, renewed war in the Middle East, inflation data that will get worse in the next print, and an economy where the consumer has been supported by depleting savings and where credit stress in areas like private credit and BDCs has been building beneath the surface. The Fed is being asked to fight inflation in a slowing economy with a renewed energy shock. That is the definition of stagflation risk, and it is the environment in which policy errors get made in both directions.

On the ceasefire itself, we have to be direct about what the July collapse and the weekend’s climbdown mean together. We wrote in our last pulse that the MOU was functioning as a floor under how bad attacks could get before commercial shipping was interrupted. That floor held through the June round of escalation and broke in July, and the reason it broke matters. The June episode ended with Iran privately backing down and both sides returning to talks within four days. The July episode has now included a second full-scale strike exchange, ballistic missile launches at US forces, Houthi involvement in the Red Sea, Saudi participation in strikes against Iran-backed groups in Iraq, and a Trump statement that talks are “a waste of time.” Trump’s Saturday post reads on its face like a return to the same pattern: threat, request from mediators, pause, negotiations. But the specific decision to cancel a strike on Iran’s energy infrastructure that had been jointly prepared with Israel is a real climbdown, and it tells us the President is more constrained by regional pressure (particularly from Gulf states whose own economies cannot handle another oil disruption) than the rhetoric suggests. The framework is under significantly more stress than it was three weeks ago, but it is also proving harder to actually kill than the July 8 declaration implied.

Oil at $85 to $90 reflects that reality. It is not the war premium of $130 in April, but it is also not the $70 floor of late June. The market has moved to a new baseline, and that baseline is higher and more volatile than anything we have seen since the war began. Watching how Brent trades in the first sessions after Trump’s weekend post will be the immediate tell. If oil pulls back sharply, the market is taking his framing seriously. If it holds above $85, the market is treating the weekend as another false start, and we should too.

The AI infrastructure and defense spending we referenced in the last pulse are worth flagging again because they are now showing up in more places. Warsh’s July 29 statement described “productivity growth and capital investment” as strong, which is not a description of an economy that is decelerating uniformly. Data centre construction, semiconductor manufacturing, and defense procurement are running hot and are inflationary in specific inputs, from electricity to industrial metals to skilled labour. These sectors do not care about the oil price, and they are contributing to the sticky core inflation Warsh cannot ignore even when energy falls. That structural pressure has not gone away, and if anything it is intensifying as the AI capex cycle deepens.

3) WHAT TO WATCH

Three things over the next three weeks will shape the picture going into September, and they are all connected.

The first is the July CPI release on August 12. This will be the first hard read of the renewed war’s impact on inflation, and given that gasoline reversed sharply in July, headline CPI will move higher than the 3.5% June print. The question is by how much. A print of 3.8% or lower keeps the Fed’s options open. A print above 4% starts to make the three hawkish dissenters at the July meeting look prescient. Core CPI is the more important number. If it moves back to 0.2% or 0.3% monthly, the argument that the June flat print was a fluke gets stronger. If core stays flat or comes in at 0.1%, Warsh has room to hold September as another hold meeting without much dissent. We think core is likely to firm up rather than moderate further, based on the shelter and services trends that have been running hot for six months.

The second is Jackson Hole on August 27-29, which is Warsh’s first appearance at the annual Fed symposium as chair. Every word will be parsed for signals about the September meeting. He has been consistent that he does not want to provide forward guidance, and his actions at the July meeting (keeping the statement short and refusing to signal the direction) suggest he intends to keep it that way. Jackson Hole is a speech rather than a statement, though, and speeches are longer. If he uses the platform to lay out a framework for how the Fed will respond to persistent inflation in the presence of energy shocks, that itself is a signal. If he simply reiterates his data-dependence, the market will assume September is genuinely open, and volatility around each data release between now and then will rise materially.

The third is whether the weekend’s climbdown produces anything real. Trump has cancelled strikes multiple times through this conflict, and the pattern is consistent: pause, negotiate, discover the substance is still far apart, resume threats. What would make this time different is a concrete deliverable in the next two weeks, either a signed protocol for reopening the strait or a formal IAEA access arrangement for the enriched uranium stockpile. Without one of those, the “perimeters of a deal” language is just another framing exercise, and we will be back to the July pattern by mid-August. The channel to watch is Muscat, where Araghchi meets Omani Foreign Minister al-Busaidi. If those meetings resume in the next ten days with concrete deliverables, the framework is repairing itself. If they do not, this weekend’s post is noise.

The thread across all of this is that we are entering a period where every data point matters more than usual. The Fed is split. The war is unresolved. The consumer is stretched. Inflation is uneven. The market is pricing an outcome somewhere between the June 17 optimism and the April 8 pessimism, but the range of plausible outcomes is wide and the tails are fatter than the current implied volatility in equities suggests. That does not mean disaster is coming. It means the margin for a policy error, an escalation error, or a data surprise is narrower than the current price of risk assets implies. Position accordingly.

Thanks for reading this 16th pulse, and see you at the next one!

AAIG is a research collective built on deep work, shared scrutiny, and quality over quantity. We treat research as a product, not content to fill a feed. By combining the expertise of multiple specialists, we find true edge where others just see noise. No hype, no daily trades. Just high-conviction ideas.

This is Asset Alliance Investment Group: Learn more on substack.com/@aaig

Disclaimer: The research and analysis published by Asset Alliance Investment Group is for informational purposes only and does not constitute financial, investment, or professional advice of any kind. All investment decisions carry risk, and recipients are strongly encouraged to seek independent, qualified financial advice before acting on any information contained herein.

So on the numbers real quick. Revenue of 171.5 trillion won, operating profit of 89.5 trillion won, a 52 percent operating margin, and almost all of it from the chip division. The phones and the TVs actually lost money this quarter, partly because they have to buy memory at the same prices as everyone else. Samsung is squeezing its own consumer business with its own chip prices. But that is simple dynamics in a scarcity market and following margins.

Following the Q&A, when the head of memory said Samsung plans to supply 60 to 70 percent of its total DRAM and NAND capacity through long-term agreements. Five-year contracts, renegotiated annually on a rolling basis. The top five global data center customers have already signed, and that is an important detail of course, because these are the buyers with the deepest pockets and the longest planning horizons. Five more customers from the AI side are in final negotiations. Customers pay substantial money up front, and about a quarter of those advance payments has already been received. Also as expected minimum price floors on mainstream products. Not only the high-end AI memory. The ordinary memory now has contractual floors under it for these customers.

This is something we have been writing about the whole year. The market refuses to give memory companies a normal multiple because it assumes the next crash is always coming. Our thesis has been that the contract structure is removing the crash, and the missing piece was always the coverage, because nobody would say how much of the business actually sits under contract. SK Hynix confirmed the architecture two days earlier, around ten customers, five-year terms, deposits, but would not give the share. Happy to see that Samsung gave the share. That is again a confirmation of the framework we track.

It is a “plan” tho not a fully completed fact yet. Part of those contracts still has to be signed. And the level of the price floors was not disclosed, so a floor can exist and still sit low. Personally i expect a good baseline as you could hear between the lines in the call.

next to that, management said the supply shortage in 2027 will be more severe than in 2026, and continues into 2028, because a new fab takes more than three years to build. Those who follow the HBM research longer know we have been positioned on 2027 as the heavily scarce year for a long time , while most of the street had 2026 as the top. the biggest memory producer in the world puts that year on the record, that open question is closed for me. Focus on 2028-2030 now even more.

And then over to pricing, because the two calls together confirmed something what led to a initial sell-off after SK hynix earnings because the street misunderstood this before earnings. Most memory reprices every quarter right now. HBM sits on yearly contracts, so its price lags for the moment. the company with the more ordinary mix should print the faster price growth this quarter. Samsung printed DRAM prices up around mid-40 percent. SK Hynix printed around 30. Logically because SK has a higher product mix of HBM. They will benefit when the repricing occurs. Two companies, same quarter, exactly the pattern. Samsung said its 2027 supply agreements are already signed, so we will see those prices come through.

On the strategy part because we focus a lot on strategy in the research.

The first point is how Samsung allocates the shortage. Management said it pretty open: the customers who commit demand and accept mutually binding conditions get the capacity, and the rest wait. And on the supply side they build the cleanrooms first and install the equipment flexibly behind actual signed demand. So the shells stand ready, but the tools only follow contracts. With this the supplier using scarcity as a selection tool, and imo it is exactly the behavior you want to see if you believe the capacity discipline holds through this cycle.

The second point is a new class of customer showing up more. Management said the frontier AI model developers, the labs themselves, are now approaching Samsung directly for multi-year supply, because they cannot secure enough capacity through the clouds anymore. Think about what that means for a second. The demand line is not only the hyperscalers now, the layer above them has started bidding too. I view that as pricing power arriving from a direction most models do not even include yet.

And the third point matters most for my own Samsung thesis, which those who follow me longer know runs through the foundry. The foundry ran at full utilization on all advanced nodes this quarter, earnings improved significantly, management expects the 2nm project wins to more than double this year, and they said on the call they are in talks with Broadcom and others. And here is the detail I like most: the base dies for Samsung's own HBM come out of Samsung's own foundry. Design, memory, foundry and packaging under one roof, sold as one package. No other supplier can offer that combination. Management called a turnaround possible in the near term, and they were honest that timing is hard to predict with custom orders. The signal I want next is the first clean profitable quarter there, because that is where the biggest part of my Samsung upside sits.

What I follow from here. First, whether the 60 to 70 percent converts from plan to signed, contract by contract. Second, any leak on 2027 contract pricing in the second half, because that is where the premium comes back or does not. Third, the HBM4 ramp. Samsung guided those sales up more than three times next quarter, and I want to see that delivered, not guided. And fourth, the thing neither company would give: how fast real wafer capacity grows. So we have to monitor that.

Good call im happy about it.

Now the whitepaper. An independent third-party systems integrator ran the test but it was not disclosed which company this was. At least it was not Accelsius running it’s own test, so that matters. The hardware was a Dell PowerEdge XE9680L which is commercially available with an eight-way NVIDIA B200 GPU server. The test swept approximately 40,000 operating points, varying GPU core power, HBM power, and CPU power across simulated AI workloads. The use of a Dell rack is interesting as well since we also called out earlier that Accelsius was part of their technology solution center where they were mentioned as their technology partner. Or it might be just an coincidence as it’s a widely available product, who knows.

It would be interesting to know if a future test is also planned with the Vera Rubin platform so we note that one down for our upcoming call with Accelsius management and technical team.

So what did the benchmark contains, first they measured the factory-installed single-phase water cooling system. Then they retrofitted the exact same server with Accelsius NeuCool two-phase cold plates and the IR150 CDU, and ran every test again. Same server. Same GPUs. Same workloads. Only the cooling changed.

Those who pay attention can see that the whitepaper only names the 8x B200 cold plates and the manifolds, and the Figure 1 caption says “all other server hardware was unchanged.” What they didn’t mention in the whitepaper but what is visible on the picture is that also included the NVLink switch cooling loop, which are shown here in the middle.

In Dell technical guide terms that’s called “Loop 3” in the B200 sled. There are five coolant loops in total, four serving the eight GPU cold plates (two plates in series per loop) and one dedicated loop serving the NVLink switch cold plates, all fed from the same rear coolant tubes as the GPU loops.

On an 8-GPU HGX baseboard like the one in the XE9680L, the GPUs don’t talk to each other directly, they communicate through NVSwitch chips that sit on the same baseboard and route NVLink traffic between all eight GPUs at full bandwidth (this is what makes the eight GPUs behave like one big accelerator for training). These switch chips are decent hardware on their own, each one burns a few hundred watts. In a liquid-cooled sled like this one, Dell gives them their own cold plates which are different parts than the GPU plates, so you can’t simply reuse a GPU evaporator there.

XE9680L selection

For the XE9680L selection we can understand the choice but it’s important to see things in perspective. Accelsius is a Dell technology partner with access to the platform, and the XE9680L is one of the most widely deployed liquid-cooled B200 servers (which is exactly the retrofit market they’re selling into). But let’s see what the benchmark actually compared. On one side, Dell’s standard cold plates plumbed two-in-series, a configuration with a downstream penalty that the whitepaper itself admits makes the second plate the thermally limiting position, served by the factory CDU.

On the other side, purpose-built NeuCool plates served by Accelsius’s own IR150 CDU. So two variables changed, not one. Also the headline of 19°C was measured at exactly the position the stock plumbing disadvantages most.

To be fair to Dell, series segments are not a Dell thing. Even NVIDIA’s reference HGX B200 layout runs GPU pairs in series with the NVSwitch cold plate somewhere in the loop. But that is exactly why the next test matters. We want to see NeuCool against the best single-phase server money can buy. For example a Supermicro DLC-2 system or a Lenovo SR780a V3 with Neptune cooling, on the same HGX B200 hardware. Or against premium merchant plates like CoolIT’s split-flow or JetCool’s microjets, at an equal CDU approach temperature and, crucially, at equal pumping power. Some of that 19°C will not survive better plumbing. The question is how much does, and we’d rather have Accelsius answer that before the market does.

The Results

So now we have placed the test in perspective we can have a look at the results, the two-phase test beats the single phase test on every metric. Thermal resistance (K/W) tells you how many degrees Celsius the chip heats up above the coolant temperature for every watt it generates. Lower is better because the cold plate extracts heat more efficiently per unit of fluid which results in less (one third) of the flow needed. Lower flow means smaller pumps, less pumping energy, and less erosion-related wear on fittings and cold plates over the life of the system.

Furthermore what stands out is that at 50°C facility water, single-phase exceeded the B200’s 84°C throttle point. NeuCool maintained 9°C of headroom below it. When a GPU hits its thermal throttle, it reduces clock speed to stay within temperature limits. Reduced clock speed means fewer tokens per second. Fewer tokens per second means less revenue per GPU per hour. In a 1,000-GPU data center running 24/7, that throttling is killing your economics. Or imagine a Neocloud saying to their client, sorry that you could not use your hourly rented GPU at full capacity because we hadn’t enough cooling capacity. Obviously this will never happen because the cooling capacity will be calculated with larger pumps or higher flow rate, but this all will impact the chiller operation hours and negatively impact the PUE of the site.

The 50°C degrees benchmark

NVIDIA has been pushing the data center industry towards warmer facility water temperatures. The reason is simple because as per the whitepaper, “every 1°C increase in facility water temperature saves approximately 4% in annual cooling energy costs”. At 50°C+ facility water, you can run “free cooling” by drawing heat out of water via heat exchangers without ever turning on a compressor-based chiller. Less chillers means lower CAPEX, less chiller operating hours means lower OPEX, a smaller carbon footprint and a lower PUE.

At 50°C facility water, the benchmark shows the single-phase system pushing the B200 past its 84°C throttle point. When a GPU throttles it drops clock speed, and reduced clocks means fewer tokens per second, which means less revenue per GPU-hour. Your $3M rack is now underperforming because the cooling can't keep up.

It’s interesting that the 50°C facility water temperature was used since the IR150 Product page itself shows “ASHRAE W45 compatible, which means it’s supporting facility water temperatures up to 45°C”. So either the W45 rating is conservative and the paperwork will catch up, or 50°C+ operation is a lab result and not yet a product.

Now, before we write “single-phase breaks at 50°C”, it doesn’t, and the whitepaper itself shows it. The system was simply never designed for 50°C water; at 45°C it does its job fine. In practice no operator would ever let their GPUs throttle either so they would fix it with bigger pumps, colder water or more chiller hours. But every one of those fixes shows up in your PUE and your OPEX, which also would mean you’re not comparing apples to apples anymore in this test.

So the right way to read the result is that the performance single-phase delivers at 45°C facility water, NeuCool matches at 54°C. At higher facility flow rates, at up to 59°C. Meanwhile NeuCool holds 9°C of headroom below the throttle point at 50°C water, growing to 14°C at higher flow rates. Those extra degrees are the entire story, because they are what moves a site from “needs chillers” to “mostly doesn’t need chillers.”

Elimination of Chillers

Chillers exist to cool facility water down to a temperature low enough for the cooling system to absorb chip heat without the chip overheating. With single-phase, that setpoint is typically 18–25°C inlet water because single-phase has high thermal resistance so it needs cold water to compensate.

Two-phase changes the equation entirely. Because NeuCool demonstrated it can keep B200 GPUs 9°C below throttle with 50°C facility water, and at higher flow rates that extends to 54–59°C, you no longer need mechanically chilled water. You just need water that’s cooler than 54–59°C. So can we eliminate chillers entirely?

In most of Northern Europe, Canada, Scandinavia, parts of the US (Pacific Northwest, Midwest), and high-altitude regions, ambient air temperatures are below 30°C for the majority of the year. A dry cooler, which is essentially a large radiator that rejects heat to outdoor air, can easily maintain facility water at 40–50°C in those climates without any refrigeration compressors running at all.

That’s what “free cooling” means, the heat rejection happens passively via the ambient environment, not via a mechanical refrigeration cycle. No compressors, no refrigerants, no chiller plant.

With two-phase in a cool climate you could eliminate:

• Chiller compressors and refrigeration circuits;

• Chiller CAPEX (typically $1–3M per MW of cooling capacity);

• Chiller energy consumption (compressors are the largest single energy draw in a conventional data center);

• Chiller maintenance contracts and refrigerant management.

Eliminating this would be a huge impact to the PUE and would therefore increase the available IT / MW. This also translates to generator sizing since these are sized to peak load, and chiller compressors represent a massive peak draw.

What stays:

• Dry coolers or cooling towers to reject heat to the environment;

• Pumps to circulate facility water;

• A small buffer capacity or backup for unusual heat waves, I would not be surprised if data centers want to keep some Chiller capacity to make sure they can bridge hot days.

The DarkNX 300MW deployment is in Ontario, Canada which is one of the better climates in the world for their cold winters and mild summers so it might not be a coincidence that Canada is where the first large-scale two-phase deployment landed. Another question to ask to Accelsius management is if the DarkNX design contains Chillers delivered by Johson Controls because we know that Johnson Controls is involved in the project too.

Adoption Curve

Essentially what this whitepaper is saying is that the operating conditions NVIDIA is actively driving the industry toward, single-phase is not just the worse choice but it’s the choice that either lets your GPUs throttle or forces you to buy your way out with more pumping power, colder water and more chiller hours. Both options cost you money, one in tokens, the other in CAPEX and PUE.

We do have to be precise here, because NVIDIA’s current GB200/NVL72 reference designs are themselves single-phase warm-water, so single-phase is clearly not disqualified today. But look at where the roadmap is going and you’ll see that warmer water every generation, and rack densities heading toward roughly 600 kW with the Kyber generation. Somewhere along that line single-phase simply runs out of thermal margin, and the operators who follow NVIDIA’s efficiency roadmap most aggressively will hit that wall first.

That reframes the adoption curve. It’s not “operators will gradually switch as they upgrade.” It’s “operators who run warm water at next-gen densities will have to switch.” That’s a transition pulled forward on a timeline that NVIDIA controls and not one that depends on cooling vendors convincing data center operators to try something new. In our experience, that second kind of transition is the slow one, the first kind is not.

However, in order to ‘switch’, the operator does need to have something to switch to. When talking about the 600kW Kyber generation, Accelsius’ biggest CDU is currently the MR250, supporting 250kW. Also the MR250 product page shows that it “Supports facility water temperatures up to 45 degrees Celsius”, while the whitepaper's whole story is the advantage at 50°C+ water. So the warm-water, chiller-less pitch isn't yet covered by their shipping at-scale product (or they just need to update the datasheet). More products should be introduced by the end of this year so we’re looking forward to see to what the capacity is and facility water temperatures are.

A hint per a AI Register article is that the “Accelsius has a 350-plus kW unit in the works, and is working on a 2.4 megawatt unit in the lab that he expects to be available sometime early next year.” This would support 4 Kyber racks although I don’t think this will happen from a redundancy point of view as one CDU failure would drop four Kyber racks. A more standard approach would be multiple CDUs on a shared manifold running N+1 at partial load, so a realistic deployment is more like five 2.4 MW units serving sixteen racks than one serving four. But this is all speculating so let’s wait till more info comes out by the end of this year.

On the retrofit side, Accelsius demonstrated this on an unmodified, commercially available server. The same Dell XE9680L that is already deployed in many data centers can be retrofitted with NeuCool cold plates without replacing the server. That creates a retrofit market on top of new deployments and it lowers the barrier for exactly the pilot projects that new cooling technology normally gets stuck on.

But one 10 kW server is not a 600 kW rack. At Kyber-class densities, the hard two-phase engineering is vapor management. Hundreds of kilowatts of refrigerant vapor returning through shared manifolds without pressure drops that raise saturation temperature and impacts the thermal margin, plus even distribution across 144 evaporators. Nobody has publicly demonstrated two-phase at full rack scale, let alone row scale so hopefully we’ll see more whitepapers coming out soon.

Takeaway

So the test is clean, NeuCool beats single-phase on every metric so two-phase liquid cooling will dominate the market, nothing to dispute right? Let me give my honest view. Nobody questions the direction of rack densities, and the physics of vaporization are not in dispute. NVIDIA’s Kyber architecture (their next-gen rack platform) likely pushes toward 600kW of power per rack, and at those densities single-phase has left the chat already. The physics don’t lie

The real question is commercial execution. I’ve worked on enough data center projects and new technology rollouts to know that the hardest part for any supplier is getting through the pilot and qualification phase with the larger players. Every connection, every fluid loop, every integration point gets checked multiple times. Pilots take 3, 6, sometimes 12+ months and they can fail on things that have nothing to do with the technology itself: leak incidents during commissioning, operator training gaps, product ramp availability, integration with facility management systems, or supply chain demands.

Some larger clients I worked for only would onboard a new technology if they had at least two suppliers to choose from so that they always had a fallback option when supplier A had a supply chain issue. For this it’s obviously much easier to have 2 / 3 Transformer suppliers than two-phased liquid cooling suppliers that both integrate in your facility perfectly. It’s all about managing risk.

The whitepaper shows good results in comparison with single phase cooling but what happens with continuous 100% utilisation, response to the millisecond power transients modern inference that workloads produce, serviceability under load, retrofit cost and downtime, and whether a retrofitted XE9680L keeps its Dell warranty (my guess is that this is why they were part of the Dell technology center).

The DarkNX 300MW deal is encouraging, but DarkNX is not a Tier-1 hyperscaler. The signal that would change the investment thesis materially is a named production deployment at AWS, Google, Microsoft, or Meta. Until then, the pipeline is real but the reference customer that triggers broad industry adoption is still missing. A reference design with NVIDIA or AMD (or both) would be a good step in the right direction and will likely have a big impact on the stock of Innventure. A reference design is essentially having NVIDIA and/or AMD saying that their products work with a two-phase liquid cooling design as well, and Accelsius is there to benefit.

Although the whitepaper gave us some valuable insights, there are some questions that are coming up which we are writing down to ask to the Accelsius technical team, here’s the top 5:

1. Independence and data. Who was the third-party systems integrator, and will you share the full test report and raw data?

2. Baseline fairness. The comparison used Dell’s stock series-loop cold plates, and each configuration ran on a different CDU. How much of the 9–14°C system-level margin comes from the cold plate physics versus the IR150 CDU, and how does NeuCool compare against a premium parallel-fed single-phase design at an equal CDU approach temperature?

3. NVIDIA relationship. Are you a qualified vendor, in qualification, or working from published OEM thermal specs? Is NeuCool on NVIDIA’s reference design for any rack configuration, and is a Vera Rubin / next-generation platform test planned?

4. Retrofit reality. What does an XE9680L retrofit involve in practice? downtime, re-plumbing, certification, cost? Was the Dell server chosen because of the technology partnership, and what does that partnership involve?

5. Scale. The test covers one 8-GPU server. What changes thermally at full-rack scale (NVL72-class: 72 GPUs, NVLink fabric, high-density power), and what facility water temperature assumptions underpin the multi-megawatt CDUs announced for before the end of 2026?

Innventure stays on hold for us until we've spoken to management, and we'll take it from there. But the list of things that could re-rate this stock just got a bit shorter and more specific. We're looking for a reference design with NVIDIA and/or AMD, the promised product announcements of the 350 kW CDU and the 2.4 MW unit reportedly in the lab. And, maybe most telling, whether the next products are actually rated and warrantied above W45. Demonstrating 50°C facility water in a benchmark is one thing; putting it on a datasheet is a commitment. We’re looking forward to those releases which could come as early as the OCP Global Summit, October 12–15 in San Jose,.

And lastly, we’re counting the day that a Tier-1 name shows up in an Accelsius press release, which would be the ultimate validation of the technology at scale.

Disclosure

We hold, or have held, a position in Innventure (NASDAQ: INV). This article reflects our opinion, is for informational purposes only, and is not investment advice. Do your own research.

Last week, I finally made a move and bought one of them.

I want to lay out the case the same way I’d walk through any position, the setup, the why-now, and what has to go right. And here’s the thing: it’s not one of the substrate makers everyone points to first. It sits a layer back.

The more I dug into who is actually furthest ahead in glass substrate, and who quietly supplies the equipment that makes their process work, the more I kept arriving at the same conclusion. Most of the market is looking at the wrong country and worrying about the wrong risk. There’s a widely repeated concern about this name that, once you trace how the ecosystem actually fits together, I think is simply overdone.

Below, I lay out why one national ecosystem sits at the center of the entire glass substrate story, why the leading player there is structurally stronger than its best-known rival, and why this one company keeps showing up at the bottom layer of it all. Plus what I expect for order flow, timing, and how I’m thinking about the setup from here.

Read more

Hi guys, welcome again to our Macro Pulse Series! Last week was the week crude oil hit $70, the May PCE confirmed inflation is still climbing in the official data, and the ceasefire saw its most violent breakdown since it was signed. Tankers were struck again, US strikes on Iran resumed, and within 48 hours both sides agreed to stop and meet again in Doha. The compression of that cycle, from weeks to days, tells us something about what kind of agreement this really is. We want to spend time on what happened, and more importantly, on what it means. Let's dive in!

1) KEY METRICS / DATA POINTS

For the key data of last week, Thursday’s PCE release was the most important domestic data point. Headline PCE rose 0.4% on the month and 4.1% annually, the highest reading since April 2023 and up from 3.8% in April. Core PCE rose 0.3% on the month and 3.4% annually, slightly above the 3.3% consensus and the highest annual core reading since October 2023. Personal spending rose 0.7% on the month against a 0.6% forecast, and personal income also rose 0.7% against a 0.4% forecast. The saving rate ticked up to 3% from 2.6% in April. Q1 GDP was also revised up to 2.1% from 1.6%.

The headline numbers here confirm what the CPI release flagged two weeks ago. Official inflation is still climbing. The energy that drove May higher has come out of the market substantially in June, with WTI roughly 38% below its May peak. Whether that shows up in materially lower June and July prints, we will know when the data is released. The lag between energy markets and official inflation prints is now the most important variable in the inflation picture, and we are reading May data while living in June reality. Those are different things.

The core number is the one that deserves the most attention, because it is not directly tied to oil. Core PCE at 3.4% annually with a 0.3% monthly print confirms that the broadening of inflation into shelter, services, and tariff-affected goods has continued. That is the part that does not unwind automatically when oil falls. It is also the part that has built quietly while the war has dominated the headlines.

On the geopolitical side, the week traced an arc compressed into days. Talks opened in Geneva on Monday June 22 without IAEA inspectors present, despite the MOU explicitly committing Iran to supervised nuclear activities. The technical team worked through Wednesday and Thursday on the inspection dispute, without resolving the political sequencing: Iran’s deputy foreign minister has tied inspection access to full sanctions termination, which is a final-deal item rather than an interim one.

Then the week broke open. Iran struck a commercial container ship in the strait on Thursday. The US responded with strikes on Iranian targets on Friday. Iran struck a supertanker carrying 2 million barrels of Qatari oil on Saturday. US Central Command launched strikes against ten Iranian military targets in and near the strait on Sunday. Trump warned on Truth Social Sunday that “the Islamic Republic of Iran will no longer exist” if escalation continued. By Monday, both sides had agreed to halt direct attacks and allow commercial vessels to transit freely. Talks were rescheduled for Doha on Tuesday. Same pattern, different week…

Oil moved in step with all of this. Brent closed at $80.57 on Friday June 19. By Friday June 26, August Brent had settled at $71.99 and August WTI at $69.23, with WTI closing below $70 for the first time since February 27. That is the day before the war began. Monday saw oil rebound modestly on the talks announcement, with WTI at $70.56 and Brent at $72.91. As of Tuesday (time of writing), Brent is trading near $74 with WTI hovering around $70. The war premium has been erased.

The EIA's most recent Short-Term Energy Outlook assumes vessel traffic ramping back to pre-conflict levels will not be complete until early 2027, which is worth keeping in mind when reading the oil price

2) AAIG’S PERSPECTIVE

The most important development of the week is the fluctuations in oil, and we want to be careful about what it actually means.

WTI below $70 with the conflict still ongoing is the market’s vote that this ceasefire holds in form even if it breaks repeatedly in substance. We think that is the right way to read it. The market is not pricing peace, but it is pricing the absence of full-scale war, partial functionality of the strait, and the assumption that weekends like the one we just saw stay contained. Those are three separate assumptions, and they all need to be right for the price to be right.

We think the market may be correct about the first two. The third is the one we are less sure of. The events of last week did not break the framework, but they showed how thin the margin is. Both sides struck each other repeatedly. Tankers were hit. The president of the United States posted a message implying he was prepared to end the existence of a sovereign nation. Within four days, both sides were back at the table. The cycle works. It also runs much closer to genuine escalation than anyone wants to acknowledge.

This is the kind of arrangement that holds until it does not, and the things that would make it not hold are not particularly difficult to imagine. Israel could escalate against Hezbollah. Iran could miscalculate one of these tit-for-tat exchanges. A weekend incident could fail to resolve in 96 hours, and commercial shipping could pause for longer than the market can comfortably digest. They are simply the kind of events that the current oil price implicitly assumes will not happen.

We have written before about the broadening of core inflation, and the May PCE confirms that this is still in motion. Core PCE at 3.4% is not consistent with a Fed that has imminent room to cut. The headline number may decelerate sharply in June and July, but the core picture is what determines policy. Services inflation, shelter, and the tariff layer continue to print firm.

This puts Warsh in a particular position heading into the July 28-29 meeting. He has been clear that he wants room to ease. The committee around him, where 17 of 18 members at the June meeting saw inflation risks tilted to the upside, is moving in the other direction. If headline data softens meaningfully in June and core moderates even slightly, he has cover to soften the dot plot’s hawkish lean. If core stays at 0.3% monthly and the broadening continues, the September meeting becomes another hold and the 2026 hike scenario stays alive. The data over the next four weeks will decide which path he is on, and his preferences will not override what the numbers say.

The IAEA situation deserves its own paragraph. Iran’s stockpile of 60-percent-enriched uranium was last verified on February 27 at 440.9 kg, which was 121 days ago (I am quick with math). The MOU commits Iran to IAEA supervision but Iran has now made inspection access conditional on full sanctions termination, which is supposed to be a final-deal deliverable. The Geneva technical team has no mandate to fix this. Grossi’s tone shifted in 48 hours last week from confident prediction to public frustration. The substantive nuclear question is not closer to resolution than it was when the MOU was signed two weeks ago. It even may be further from it.

For asset prices, the picture comes down to this. Equity markets and oil are pricing normalisation. Bond markets are pricing inflation persistence. Both can be right for a while. The reconciliation comes when either oil moves materially in one direction or another, or the inflation data forces the Fed’s hand. We are entering a period where headline data may flatter the picture, core data may continue to tell a different story, and the geopolitical situation is genuinely unstable but contained. That is a difficult environment to position for, because the obvious trade (faded oil, lower rates, equities up) assumes the floor holds. We are less confident in the floor than the market currently is.

3) WHAT TO WATCH

Three things will shape July, and they are all variables that are connected.

The first is whether the Doha talks produce anything substantive on the nuclear question. The technical team in Geneva has been working on inspection access. The political-level conversation in Doha needs to bridge Iran’s sequencing demand with the US insistence that IAEA access is an interim deliverable, not a final one. Without that bridge, the MOU’s commitment to IAEA supervision is hollow, and any final deal becomes much harder to reach. If Doha ends with another vague statements about progress without concrete movement on inspections, the agreement is functionally an oil-for-ceasefire arrangement and nothing more.

The second is whether the oil floor at $70 holds through July. We have argued the market is voting on the absence of war, not the presence of peace. That vote is conditional on no major escalation that interrupts commercial shipping for more than a few days at a time. The base case is that tanker transit counts recover gradually from this past weekend’s slowdown. If they stall below pre-conflict levels and stay there, the inventory drawdown continues even with sanctions relief on Iranian exports, and the floor becomes harder to defend into August.

The third is the June CPI release on Tuesday July 15 and the June PCE on Thursday July 31. These two prints will set the tone for the July 28-29 FOMC meeting and the September meeting. The headline numbers will likely show energy relief feeding through. The core numbers are what matter for the Fed’s actual reaction function. If June core CPI prints at 0.2% monthly with annual core PCE moderating to 3.3% or below, Warsh has the data he needs to begin softening his tone. If core stays at 0.3% monthly and PCE holds at 3.4%, the September meeting will be another hold and the hike scenario stays alive.

The thread this week is that the market has priced an outcome more stable than the events currently support, and we have to be honest about why. The headline data may flatter the picture in the coming months. The structural problems do not move with it. The IAEA stalemate, the unresolved nuclear question, the compressed escalation cycles, the broadening of core inflation, none of these is closer to resolution than they were two weeks ago. We are not arguing that something bad is imminent. We are arguing that current asset prices reflect more certainty about the path forward than the situation actually shows us. The next escalation will come, and when it does, the floor will be tested. The probability that it holds is real but not as high as $70 oil and continued equity strength suggest.

Thanks for reading this 15th pulse, and see you at the next one!

AAIG is a research collective built on deep work, shared scrutiny, and quality over quantity. We treat research as a product, not content to fill a feed. By combining the expertise of multiple specialists, we find true edge where others just see noise. No hype, no daily trades. Just high-conviction ideas.

This is Asset Alliance Investment Group: Learn more on substack.com/@aaig

Disclaimer: The research and analysis published by Asset Alliance Investment Group is for informational purposes only and does not constitute financial, investment, or professional advice of any kind. All investment decisions carry risk, and recipients are strongly encouraged to seek independent, qualified financial advice before acting on any information contained herein.

There is a road in the south of the Netherlands where the traffic lights seem to wait for you. Drive it at the right speed and each signal turns green just as you reach it, one after another, the whole corridor opening ahead of you. A small board on the roadside tells you the number to hold. Fifty-three. Nothing else. No screen full of data, no app, no alert. Just a number on a sign, and somewhere behind it a system that has already counted the cars ahead of you, measured the gaps between them, worked out the timing of every light on the route, and decided that fifty-three is the speed that makes the city dissolve quietly beneath your wheels.

I keep coming back to that road. Not because it is impressive technology, exactly. It is the opposite of impressive in any way you can point at. It is a sign with a number on it. What stays with me is the strange shape of the thing underneath, because the more I look at it the more it seems to describe something much larger that is happening all around us, mostly unseen.

Think about what that road is doing. An enormous amount of computation went into producing the smallest possible result. All of the intelligence stayed hidden, and the only thing that reached me, the person in the seat, was a single number I could choose to follow. That inversion is worth sitting with. The cleverness runs deep and out of sight, and what surfaces is almost nothing — one quiet instruction. I have started to think of this as ambient AI: intelligence dissolved into the physical world, invisible while it works, surfacing only as the simplest thing it can say to you. Not something you open or talk to. Something you move through without noticing it is there.

We have spent three years rightly transfixed by the other kind of AI, the kind that talks. The models that write and reason and code, the agents that now spin up a working website in the time it takes to describe one, the systems that seem to do something new and astonishing every few weeks. That attention is not misplaced. That really is where the most dramatic change is happening, and it is what pulled most of us into this whole subject to begin with. I am as caught by it as anyone.

But all that brightness has a way of stepping over a second thing happening at the same time. Quieter. Slower. Far less spectacular to watch. And, I have come to think, no smaller in what it will eventually mean. While we look up at the frontier, the ground beneath us is starting to change.

## The century of dumb concrete is ending

For about a hundred years, infrastructure has been a fixed thing. You poured the concrete, wired the grid, hung the traffic light, laid the rail, and from that day on it ran the same rigid program until a crew came to repair it. A traffic signal in 1970 and a traffic signal in 2020 did fundamentally the same thing. Each one counted down a timer it had been handed at installation, indifferent to whether the road in front of it was jammed solid or completely empty. The grid pushed power in one direction and hoped supply matched demand. The pumps ran on a schedule. The trains followed a timetable drawn up months in advance. None of it could see anything. None of it could think. Infrastructure was a finished object, complete the day it opened, and slowly decaying from that moment forward.

That is the part that is now ending, and it is ending faster than most people have noticed.

What is taking its place is infrastructure that behaves more like a living system. It senses its own state as things happen, works out where they are heading, decides what would be better, and acts. Then it does the whole loop again, continuously. And the thing I find genuinely striking, the thing that makes this one story rather than a dozen unrelated ones, is that it is the same loop everywhere you look. The traffic light learning to wait for you and the power grid balancing itself thousands of times a second are not two separate trends. They are the same idea.

Once you see the shape, your mind automatically starts seeing it everywhere. Here in the Netherlands, the agencies that have kept a country alive below sea level for centuries are moving to models that compute a flood scenario in seconds where the old simulations took hours, closing barriers on a forecast before the water has even begun to rise. The same loop is quietly turning up in the rail network routing its own trains, in the port reading tides and currents to slot a ship into a waiting berth, in the district heating that warms a whole city by weighing tomorrow's weather against today's demand. The engineers who built that last one describe their goal as doing all of it without the residents ever noticing, and that phrase has stayed with me, because it is really the whole spirit of the thing. The highest compliment this kind of intelligence can earn is that you forget it was ever there.

The biggest challenge in that loop shows up most clearly in electricity. The data centres behind the loud AI are consuming power at a pace that is bending the grid out of shape. You cannot answer that with new power stations alone. So instead, the grid is being taught to think. Hundreds of thousands of home batteries and water heaters and thermostats are being woven together into what amounts to a single power plant with no chimney, one that exists only as coordination, that an algorithm can call on in concert at the moment it is needed. I already shared my thoughts on this back in October, when I wrote about the AI energy bottleneck and the early, mid, and end stages of that market.

It is worth pausing on what this buys, because the gains are not small. We are pulling capacity out of infrastructure we already own, simply by letting it see and decide, without pouring any new concrete at all.

And what makes this the moment, rather than some moment ten years ago, is that the thinking is getting cheaper. The cost of a given amount of AI capability has been falling, and is still falling. When the thing that makes a system smart keeps getting cheaper like that, you stop reserving it for the single most important asset and start putting it into all of them. Every light. Every pump. Every transformer. Every berth and every train. The economics that kept infrastructure dumb for a century have simply turned over.

## It arrives unevenly, and that is telling in itself

The change is happening, but of course not everywhere at once. It arrives first where the ground is soft.

You can already see what it buys where it has taken hold. When the city of Hangzhou let a system coordinate more than a thousand of its traffic signals as one network, rather than a thousand stubborn separate timers, the city fell from the fifth most congested in China to the fifty-seventh. Ambulances began reaching emergencies in roughly half the time they used to, and the pressure came off the roads without a single new lane being laid. Nothing was built to achieve it. The same city, suddenly able to think.

In the Gulf, they are trying something bolder still, skipping the slow retrofit entirely and building cognitive cities up out of bare desert, with the coordination designed in from the first drawing rather than bolted on later. Whether the most extravagant versions of that vision ever get fully built is a fair question. But the approach itself, starting from nothing with no legacy to tear out, is the fastest path there is, and it shows you the shape of what a city built to be orchestrated wants to become.

The change moves fastest where decisions are centralised, where the build is new, where a society is comfortable letting a system act on its behalf. The places best suited to deploy it quickly are not always the same places that invented the pieces.

## A day inside the quieter world

Let me try to describe an ordinary morning a few years out. To really understand the size of it, you want to visualise it.

You wake in a home that has been quietly managing itself through the night. The heat pump did most of its work in the small hours, when the wind was strong and power was nearly free, banking warmth in the mass of the building so that the morning costs almost nothing. You never touched a thermostat. You never thought about any of it. The house optimised itself around you while you slept, and the only trace of it is a room that happens to be exactly the right temperature and a bill that is smaller than it has any right to be.

You get into the car, and the route that loads is not the shortest one, and not even the fastest one taken on its own. It is the route the network has chosen for you knowing where every other car leaving right now is also trying to go, so that all of you move better together than any of you could alone. You reach the corridor and the lights open ahead of you in sequence. Fifty-three again. The city dissolves. Somewhere a system has folded your single trip into the movement of ten thousand others and found the arrangement in which nobody has to fight anyone, and none of that reaches you. What reaches you is only that the drive was easy.

The car a little way ahead is an ambulance, and three intersections before it arrives the lights have already cleared its path, because the network saw it coming and bent the city gently around it. Minutes saved. Now and then, a life.

Your car charges itself in the early afternoon instead of the evening, not because you scheduled it but because the grid asked, offering a few cents and a cleaner conscience to move your demand to the hour when the sun is highest, and the car simply accepted on your behalf. Multiply that small unseen agreement by a million cars and you have flattened a curve that used to require a fleet of gas plants standing ready. The plants that never had to be built are, in a way, the most elegant infrastructure of all. They exist only as an absence, as something quietly made unnecessary.

And below all of it, out of sight, the heavy machinery of the world is breathing in the same rhythm. The water network reroutes around a building pressure spike before it becomes a burst main in the street. A port two hundred kilometres away slots an arriving ship into a berth that opens at exactly the hour it is needed, so the ship never sits idling offshore burning fuel. A freight train you will never ride finds a path through the network that appeared on no timetable. None of this asks for your attention, and that is exactly the point.

This is the texture of the world that is coming. Friction quietly dissolving in a thousand small places at once. Every wait at a red light, every spike in a bill, every flooded underpass, every ship idling outside a crowded port is, when you look closely, a coordination failure, a place where infrastructure that could not see far enough simply failed to arrange things well. And one by one, those places are starting to be seen. The remarkable part is how little of it you will ever perceive. The intelligence makes its progress precisely by disappearing. You will not marvel at it. You will just notice, on some unremarkable Tuesday, that the day put up a little less resistance than it used to.

That is the thing I cannot stop turning over. After a century of building infrastructure that could only ever decay from the day we finished it, we are starting to build infrastructure that gets a little better every day it runs. It does not announce itself. It will never trend. And it may quietly touch more of ordinary life than the louder revolution we are all, quite reasonably, watching instead.

Which brings me to the part that is harder to see from the outside, and worth saying plainly.

When you visualise this world and think about where the value is captured, the instinct is to look at the intelligence itself, the brains of the thing.

A part of the value seems to settle somewhere far less glamorous. It settles in the layer that sits on top of all the messy physical hardware, the pipes and poles and meters and controllers, the layer that the intelligence has to run through in order to touch the real world at all. To own that layer you have to own the unglamorous part, the integration into thousands of municipal contracts and utility certifications and regulatory approvals, the slow grinding work no clever model can simply leap over. The intelligence has become the easy part. Being the thing the intelligence has to plug into, the part that is certified and embedded and depended upon and paid month after month, is the hard part, and the hard part is where the value protects itself.

It is worth being concrete about how this works, because there is a quiet asymmetry in it. The people who own these systems, a city, a water authority, a power company, sit on the hardware and the data, but the last thing most of them want is to become a software company themselves. They would much rather someone else handled the thinking. So whoever steps in to be that thinking layer ends up in an unusual position. They become the part nobody especially wanted to own, that everybody nonetheless has to run through. And once a city's intersections or a utility's grid are breathing through your system, you do not get unplugged on a whim. You have become load-bearing.

That changes what such a company is worth, in a way that is easy to miss from the outside. It begins life selling a thing, a box, a sensor, a contract, the sort of sale that competitors copy and that prices grind steadily toward the floor. But if it becomes the layer the system genuinely runs on, it stops being paid for a product and starts being paid simply to keep things running, month after month and year after year, because the alternative is the lights going out. That is the same shift I have watched reshape other corners of technology, where something that was sold as a cheap commodity quietly turns into something priced like infrastructure, like a bridge or a power station. When that turn happens to a company, its position does not just improve. It changes category.

A good example to point to is Kraken, the software arm that grew up inside Britain's Octopus Energy. Over the past few years it quietly became the operating system running the billing, the grid services, the customer machinery for utilities in a long list of countries, contracted now to more than seventy million accounts and processing fifteen billion data points a day, with over half a billion dollars of contracted recurring revenue that has roughly quadrupled in three years. As it spins out toward independence in 2026 it carries a valuation close to nine billion dollars before it has even reached the public market, and its chief executive talks openly about reaching a billion lives within a decade. That is the scale this layer can reach when a company truly owns it. Not a feature sitting inside something larger, but the nervous system of an entire industry, priced like the critical infrastructure it has quietly become.

You can watch the same playbook one layer over. NoTraffic is doing to the traffic signal something close to what Kraken did to the meter, turning intersections into software it runs centrally, live now in more than four hundred agencies and pushing toward one in every ten across North America, still private and still early. Miovision is working the same intersections from a different angle. These are the companies quietly settling into the nervous-system layer of physical infrastructure before most people have a name for the category at all.

And here is the detail I find most telling. The public market keeps losing access to this layer almost as fast as it appears. The pure plays that do list tend to be bought and taken private before ordinary investors can compound alongside them. Iteris was absorbed by an Italian group. Yunex was taken for around a billion. Q-Free and Volue and others were quietly delisted, the grid-software specialists folded into larger conglomerates and private buyers. The reason is not mysterious, and it is really the same point the whole essay keeps circling. A private buyer can look at one of these companies and see clearly what it is, contracted revenue, deep switching costs, a position embedded in critical infrastructure, and can value all of that precisely. The public market, meanwhile, often still files the same company under boring industrials and declines to pay up for it. So the private buyer quietly takes it, arbitraging a gap the public has not yet woken up to. The smart, patient money is not distracted by the noise at the frontier. It is down here, buying the quiet layer, while most attention stays up where the lights are brightest.

None of this is going to be smooth, and it would be dishonest to pretend otherwise. The road from a working demonstration to a deployed system runs straight through procurement budgets and liability arguments and data-governance fights and the slow grinding politics of public institutions. And a system that acts on the world can act wrongly, with the stakes rising the moment infrastructure stops merely reporting and starts deciding. All of that is real, and all of it will slow things down.

But none of it changes the direction of travel. It only changes the timing. The cost of intelligence keeps falling, the load on our infrastructure keeps climbing, and the only real way through is to let the concrete start to think. The loud revolution is teaching machines to speak, and it deserves the attention it gets. The quieter one is teaching the world to listen to itself, and it is happening so close underneath us, so deliberately and so without spectacle, that it is easy to look straight past it on the way to something brighter. On some ordinary morning a few years from now you will move through a day that has quietly arranged itself around you, the lights opening, the power flowing, the water and the rail and the steel all breathing in a rhythm you never see. You will not notice a single line of it. That is exactly why it is being underestimated.

So let me say plainly what I have come to believe, because it is the thing I am most convinced of. There is going to be a layer in all of this that captures an enormous share of the value. It is the unglamorous operating layer that the physical world comes to depend on and cannot easily unplug, the part that gets paid simply to keep things running. That is where the durable value in ambient AI is going to settle, and I do not think it is close. Kraken is what it looks like early, near nine billion dollars before it has even gone public, on revenue that is contracted, recurring, and embedded. The traffic and grid and water versions of that company are being built right now, most of them still private, several of them being quietly bought before the public ever gets a clean look. A part of the smart money is already down here, in the plumbing, while the rest of the attention stays up at the frontier where the lights are brightest.

That is the part I would not look away from. The revolution everyone is watching is real. But the one that ends up running the world you wake into each morning may be the quiet one.

In this part, we bring it all together into a framework you can actually use. We lay out the market model and how it links to specific stocks, what to monitor going forward, the catalysts to watch through 2028, and the risks that could break the thesis, before closing with our overall conclusion.

• 11. Market Model and Stock Linkage

• 12. Monitoring

• 13. Catalyst Calendar 2026 to 2028

• 14. Risk Section

• 15. Conclusion

Read more

1) KEY METRICS / DATA POINTS

The MOU signing on Tuesday was the central event of the week. New ceasefire conditions were agreed on June 12, and the presidents of the US and Iran signed a memorandum of understanding on June 17 that established a 60-day extension of the ceasefire to negotiate the final terms. Following the signing, the US lifted its blockade of Iranian ports and is allowing Iran to sell its oil freely. The deal calls for Iran’s assets to be unfrozen, though the pace is not yet clear, and it outlines a $300 billion fund for postwar reconstruction conditional on a final agreement being reached.

The operational picture in the strait shifted quickly. Vance said on Thursday June 18 that tankers carrying more than 12 million barrels crossed the strait overnight, and that the Iranians did not shoot at any ships for the second consecutive night. Ships actually moving without being targeted is the threshold that has been missing for four months. If that pattern holds through July, the normalisation of global oil markets is finally underway in practice rather than only on paper.

The May CPI release on Wednesday June 10 set the backdrop for the Fed meeting. Headline CPI rose 0.5% on the month, putting the annual rate at 4.2%, the highest reading since April 2023 and up from 3.8% in April. Energy prices jumped 3.9% on the month, with the 12-month increase reaching 23.5%. Core CPI, however, rose just 0.2% on the month, half the April pace, with the annual rate at 2.9%. The headline ran hot on energy, the core ran softer than consensus. That mixed picture gave Warsh a usable hand a week later.

The FOMC decision on Wednesday June 17 broke the pattern of the past several meetings. The committee voted 12-0 to hold the federal funds rate at 3.50% to 3.75%. Nine of the 18 voting members project a rate hike before the end of 2026, with six projecting two 25-basis-point hikes. The median projection for year-end 2026 rose to 3.8% from 3.4% in March, and 17 of 18 officials judged the risks to their inflation forecast to be weighted to the upside. Warsh did not submit his own projection for the dot plot, and the FOMC statement was cut to 130 words from 341 in April, removing the prior easing-bias language entirely.

The diplomatic picture broke open again at the end of the week. Talks between the US and Iran were called off on Friday June 19 after intense fighting between Israel and Hezbollah in southern Lebanon. Iranian officials did not travel to Switzerland as planned, insisting that the fighting in Lebanon must stop before talks can take place. Israel and Hezbollah later agreed to renew their ceasefire, mediated by Qatar, the US, and Iran.

Oil moved lower across most of the week then partially recovered on Friday. Brent crude futures closed Friday at $80.57, with WTI at $77.54. Compared with Brent around $97 a month ago and the conflict peak above $130, the move to $80 reflects real progress. The Friday bounce shows the market is not yet convinced this holds.

2) AAIG’S PERSPECTIVE

The MOU getting signed matters, and the strait moving 12 million barrels per night matters more than any of the diplomatic statements made over the past four months. The threshold has changed from announcements about opening the strait to actual tankers passing through without being fired on. If that pattern holds through July, oil markets normalise in the hard data and the inflation pressure we have written about in recent pulses begins to ease.

Friday made clear that the structural risks have not gone away. The same problem that broke earlier rounds of negotiation, Israel striking Hezbollah while the US tries to negotiate with Iran on Lebanon, has now broken the first scheduled round of post-MOU talks. Iran’s position is consistent: it will not engage on nuclear questions while Israel continues operations in Lebanon. Israel’s position is also consistent: it will not stop operations against Hezbollah to accommodate a US-Iran deal. The framework signed on June 17 does not resolve that incompatibility. It defers it.

The most likely pattern over the next 60 days is probably the one we saw this week in miniature. The operational ceasefire holds for tankers, oil keeps flowing, but the diplomatic process around nuclear talks moves unevenly, with each Israel-Hezbollah flare-up causing Iran to delay or cancel a scheduled meeting. That can continue for weeks before it threatens the broader deal. The risk is that one of those flare-ups eventually triggers an Iranian response in the strait, and the operational picture deteriorates quickly.

On the Fed, Warsh’s first meeting was a clearer statement of intent than markets had expected. The dot plot flipping from one cut in 2026 to potentially two hikes, with 17 of 18 officials seeing inflation risks tilted to the upside, is a meaningful hawkish shift. The headline rate did not move, but the framework for thinking about where rates go next did. Warsh’s refusal to submit his own dot, and the much shorter statement, both fit his long-held criticisms of Fed forward guidance, and they also remove some of the guidance markets have relied on. Investors now have to make decisions about rate paths with less anchored signalling than at any point in the past decade.

There is a tension inside the meeting between Warsh’s personal disposition and the data. Warsh has called AI “structurally disinflationary” and told the Senate in April that “inflation is a choice.” He has dismissed core PCE as “a rough swag” and favoured trimmed-mean measures that have run closer to 2% than to the headline. That is the profile of a chair who wants room to cut. The data is not letting him, and the committee around him is moving in the opposite direction. Holding the line on a hawkish projection in his first meeting, rather than fighting his colleagues on it, signals that even a chair who wants to ease cannot do it against 4.2% headline CPI and a labour market that has not cracked.

The softer core CPI reading is the data point that gave Warsh some cover, but one month of softer core does not establish a trend. Energy is still pulling headline higher. The pass-through into services, shelter, and apparel that we wrote about after the April release moderated in May, but whether that moderation continues or May was noise around a still-elevated trend remains an open question. June and July data will answer it. If core CPI re-accelerates, Warsh’s hand becomes harder to play, and the nine dots projecting a hike start to look directional rather than precautionary.

On oil, the move to $80 Brent reflects the market beginning to price normalisation, though normalisation here does not mean a return to pre-war levels. Refinery damage, depleted inventories, and the lingering insurance premium will probably keep prices structurally higher for months. Downside scenarios that get to $70 or below require both a durable ceasefire and a rapid recovery of physical infrastructure, neither of which is in place yet. The base case from here is probably Brent in the $75 to $85 range through summer, with upside risks dominating if the Lebanon situation escalates again.

3) WHAT TO WATCH

Three things will shape the next two weeks, and they are tightly connected.

The first is whether the Geneva talks get rescheduled and actually held. The 60-day clock on the MOU started on June 17, which means the substantive nuclear negotiations have to make meaningful progress by mid-August. Every week lost to Israel-Hezbollah escalation is a week not spent on the harder questions. The signal to watch is whether Vance and the Iranian delegation actually meet face-to-face within two weeks. If they do, the framework is alive. If three weeks pass without a meeting, the process is in trouble.

The second is tanker transit counts through the strait. The Vance figure of 12 million barrels per night is the operational benchmark to measure against. If that pace holds, oil keeps drifting lower and inflation gets the relief it needs to break out of the trend we have flagged in previous pulses. If transit counts fall sharply in response to another Lebanon flare-up, the deal’s credibility erodes and oil moves back toward $90. Daily and weekly transit data has effectively become the leading economic indicator for the rest of the year.

The third is the June CPI release on Tuesday July 14, which is the first reading that will capture meaningful deal-related energy price relief if it holds. A headline reading materially below 4.2% with core staying around 2.9% would make the Fed’s hawkish lean from this meeting look less necessary, and Warsh would get room to soften his tone. A headline that stays elevated and a core that re-accelerates would make the dot plot’s hike projections feel directional rather than precautionary, and bond markets would reprice accordingly.

The thread this week is that the headline progress is real, the operational signs are good, but the structural fault line running through Israel-Hezbollah has not closed. We have a signed deal, an open strait, ships actually moving, and a Fed that has stopped pretending cuts are coming. We also have a diplomatic process that got postponed within four days of being signed because Israel cannot stop striking Lebanon. The next 60 days will determine whether the MOU consolidates into a real settlement or ends up as another framework that worked for as long as nobody pushed too hard against it.

Thanks for reading this 14th pulse, and see you at the next one!

AAIG is a research collective built on deep work, shared scrutiny, and quality over quantity. We treat research as a product, not content to fill a feed. By combining the expertise of multiple specialists, we find true edge where others just see noise. No hype, no daily trades. Just high-conviction ideas.

This is Asset Alliance Investment Group: Learn more on substack.com/@aaig

Disclaimer: The research and analysis published by Asset Alliance Investment Group is for informational purposes only and does not constitute financial, investment, or professional advice of any kind. All investment decisions carry risk, and recipients are strongly encouraged to seek independent, qualified financial advice before acting on any information contained herein.

From here, we move from understanding the sector to acting on it. In this part, we get to the heart of it: who is best positioned to win at each layer, where the real money is already moving, how the most realistic adoption paths play out, and what could delay the whole thing.

• 6.6 Best Positioned by Layer

• 7. Capital Movements at Companies Tied to Glass Substrates

• 8. Realistic Packaging Integration Paths

• 9. Adoption Path: Scenarios and Timeline

• 10. Launch Delay Drivers

Let’s start with a quick recap overview of the different layers.

Read more