“They are just flying empty airplanes around.”
- Son
NVIDIA is the North Star, Orion, the whole Milky Way.
NVIDIA is benefitting from strong demand, but is selling into a concentrated set of buyers whose own demand is being distorted by a training and benchmarking phase that will not last. That distorted demand is working like a bullwhip into NVIDIA’s own supply chain through custom supply commitments as well as downstream into data-center financing. Looming over it all is the bezzle, which once seen, cannot be unseen, and once revealed, does not exist.
Fortunately, its Luminous, its Grace reported big positive headline numbers this week, so a fresh 10-Q is upon us. I love the smell of 10-Qs in the morning.
Some are watching inventory, I know, but inventory does not stand out. The inventory breakdown makes sense. Raw materials inventory surged in Q1 FY27 but is consistent with the Blackwell Ultra ramp as well as the Q4 FY25 Blackwell ramp pre-positioning of inventory.
More interesting are NVIDIA’s customers. Customer concentration is off the charts.
NVIDIA’s biggest customer is likely Microsoft, whose CFO said in January that Azure capacity constraints would persist at least through June of 2026. If Microsoft cuts capital expenditure on NVDA chips 20%, that is a 4.2% revenue hit to NVIDIA.
Cisco never had even one 10% customer. Cisco needed a more correlated pull-back to be hurt. NVIDIA is hurt badly if just one customer pulls back, or even if that one customer does not grow its orders as fast as NVIDIA needs it to.
Concentrated Mafia Don customers, concentrated bespoke supply chain. Oh boy.
Neither Cisco nor Sun Microsystems had it this bad, and Sun completely went down. I am seeing both in NVIDIA, as well as shades of historic DRAM cycles. This is all very dense stuff, as below.
Customer concentration drove DRAM cycles to their violent natures. DRAM cycles became less customer concentrated, less theme-correlated over the last decade or two, although today we are seeing a 1990s reprise on demand from a few big customers/one big theme again. I am short the SOX (the Philadelphia Semiconductor Index) for this reason.
Cisco wrote down half its purchase commitments, which it bought from suppliers who had many other customers. Cisco bought components that were in general, fungible.
This is not what NVIDIA is doing. NVIDIA is committing to custom lines at TSMC and funding them. TSMC cannot make NVIDIA’s chips with its normal lines. That is a custom non-fungible supply chain. Not only geographically risky, but likely to result in significant losses should NVIDIA’s business slow or heaven forbid suffer a decline.
With a market capitalization over $5 trillion, $182 billion in forward purchase commitment – $119 billion with one customer – does not seem significant.
In fact, the $182 billion is greater than NVIDIA’s prodigious annual operating cash flow.
None of this has been a problem as yet. Microsoft has been buying chips faster than it can deploy them. These chips, so critically, are construction-in-progress (CIP) and do not depreciate on the balance sheet or cost Microsoft anything at all until they are put into service.
When Microsoft buys those chips however, NVIDIA books the sale out of its inventory. NVIDIA’s inventory numbers look good, as I already established.
NVIDIA gets a sale, an accounts receivable that will turn into receipt of cash and hence positive cash flow.
I spy a wrinkle though in the accounts receivable pattern.
Microsoft/Customer A inflected up as a percentage of NVIDIA’s Accounts Receivable as it inflected decidedly down as a percentage of NVIDIA’s Revenue during the 1st quarter. Microsoft’s receivables balance with NVIDIA stands at about $12.2 billion now, roughly the size of NVIDIA’s entire all-comers receivables balance in 2024.
Yes, the whole business has grown. The last few years, revenues grew 4.9x. Total Accounts Receivable grew 4.9x. Customer Accounts Receivable balance grew 13.4x. Microsoft swooped in large.
This is not arguing growth. This is about a zig where there were only zags. Neither customer B nor Customer C zigged. Only Customer A. Microsoft, which has been making noise and news for a couple years now about not wanting to be in the data center business – bare metal, they said.
Microsoft is throwing popcorn all over the trail, with neither a thought to feed the fauna.
In 2024, Microsoft’s CEO Satya Nadella said, “you may actually have a bunch of chips in inventory that I can’t plug in.”
In April 2025, Microsoft froze 1.5GW worth of data center projects, backing off on leases. This was discussed in Blessed Fraud.
Microsoft’s fiscal year ends June 2026, and in January of this year, Microsoft said it expected capacity constraints through June. Over the next several years, Microsoft says it will continue to scale up its own Maia chip.
Then on May 14th, Microsoft announced it was winding down Claude Code and forcing all employees to its internal GitHub Copilot by June 30th. That is some fast compression. Claude Code was just introduced to Microsoft engineers December 2025.
This brings up two scenarios.
One, Microsoft pulled forward inventory it does not really need. It takes delivery even though builds are slowing and power shortages are biting so as to keep its priority spot as a customer of NVIDIA for the next generation chip. Just slot them into CIP (drop them on the warehouse floor) where they do no harm, depreciating financially not a whit while physically depreciating rapidly. Plus Microsoft has what appears to be 60-65 day terms with NVIDIA, so no cash leaves Microsoft’s campus for a couple months.
Two, NVIDIA pushes inventory forward to make quarterly numbers pop, beating the Street’s whispered expectations. NVIDIA knows Microsoft is already warehousing its chips, and is willing to absorb chips. Microsoft may have even cooperated in this way before. Besides, Microsoft can take that inventory into CIP with no consequences to its own earnings or reported expenses.
Meanwhile, S&P Global sees widespread grid and power shortages for data centers, as well as a 19 GW shortage of power (40% of need) for data centers by 2028.
Everyone is finding data centers to be the rate limiting step. And all those NVIDIA chips in CIP just sit there waiting.
In either case, this is a set up on the down low for a wicked bullwhip. I have discussed the bullwhip effect in prior articles. The bullwhip played out remarkably on supply chains during the COVID lockdown, and the bullwhip is the reason Cisco wrote down half of its forward supply commitments in 2001.
This is not quite a smoking gun, but more of a finding of a finger on the trigger. Watch the next few quarters, especially as the data center financing story evolves.
Meanwhile, bulls cannot stop talking about AI demand. How insatiable it is. How insane it is.
I agree, it is insane. It is the bezzle. It is not coming from a useful, repeatable steady state, and not from an ever upsloping growth curve, but rather from a temporary, social, organizational phase as companies of almost every stripe fight for their competitive advantage – their own AI stack.
To do this, companies all over are benchmarking, creating and collecting traces, training wrappers and gathering failure nodes as fast as humanly possible.
This crazy, rushed, temporary phase has a name.
The full article is available on Substack.
Please refer to the link below.
https://michaeljburry.substack.com/p/the-heretics-guide-to-ais-stars-part?r=4repfn
Sections Remaining:
- Tokenmaxxing
- The Bezzle
- The Tokenmaxxing Pyramid
- Jevons Paradox
- Where is the Funding Coming From?
- Life Insurance to the Rescue?
- The FHLB?
- The Closed Loop
