The AI Boom Has Reached the Snack Aisle
What HBM Chips reveal about the infrastructure powering AI
Cypher Capital | Kim Wong | September 2026
A memory chip becomes a snack. SK hynix and 7-Eleven's HBM Chips in Korea.
On a recent trip to Korea, I found one of the most important components in the AI boom sitting in a 7-Eleven snack aisle. It came in a bright yellow bag carrying the SK hynix logo. It was called HBM Chips.
The joke only works if you know why those three letters matter. HBM stands for high bandwidth memory, the stacked memory that sits beside advanced GPUs and moves data fast enough for AI models to train and run. SK hynix is one of the companies at the centre of that supply chain.
On the packet, HBM had a second meaning: honey banana mat, with mat meaning flavour in Korean. Inside were square corn snacks shaped to look like semiconductor chips. SK hynix and 7-Eleven launched the product to make a specialist technology feel familiar. The first production runs sold out, and sales jumped again after Jensen Huang opened a packet and shared it during a visit to Seoul.
It would be easy to dismiss this as clever marketing. I think it says something more important. AI has moved beyond research labs, earnings calls and investment conferences. In Korea, one of the most complex components in the AI stack has become a convenience store snack, a cartoon character and part of popular culture. The infrastructure behind AI is becoming part of daily life before most people fully understand what it does.
That is why I do not think the AI boom is imaginary. The technology is real, demand is real and the buildout is already reshaping companies, capital markets and daily life. But real technologies can still be financed badly. The investment question is not whether AI matters. It is where durable value sits in the stack, who can finance the build and who is taking more risk than the market realises.
To understand why that distinction matters, it helps to go back to the last technology boom that was broadly right.
Back in 1998, one number was everywhere in Washington and on Wall Street: internet traffic was doubling every hundred days. A US Commerce Department report repeated it. So did analysts, journalists and executives raising money to lay fibre optic cable. If it was true, the world needed almost unlimited bandwidth. The people who owned the pipes would own the future.
The number was wrong. Andrew Odlyzko, a mathematician at AT&T Labs, went looking for the data and could not substantiate it. His work suggested traffic was closer to doubling every year. That was still extraordinary, but nowhere near the claim. He traced the hundred day figure back to marketing by UUNET, a WorldCom subsidiary that clearly benefited from convincing investors that bandwidth demand had no limit.
Odlyzko published the work, but it made little difference. The capital kept coming. In early 1999, Alan Greenspan compared internet investing to buying a lottery ticket. Investors knew they were overpaying, but they hoped a few winners would justify it. By 2002, much of the fibre built during the boom was still dark and unused. The demand eventually arrived, just not in time for the people who financed it.
The uncomfortable pattern
That is what should make investors uncomfortable about AI today. The internet story was broadly right. It changed commerce, bandwidth demand exploded and the fiber built during the boom became essential infrastructure for streaming, cloud computing and now AI. The people who said the technology would reshape the economy were right.
They still lost their money.
This pattern keeps repeating. Railways transformed Britain. The internet changed commerce. Housing remained a useful long term asset. Investors were not ruined because the technology or the asset was worthless. They were ruined because the financing assumed capital would always be available. The break came when credit stopped reaching the people who needed it most.
The Railway King
Britain went mad for railways in the 1840s. At the centre was George Hudson, a York linen draper who inherited money and became one of the country's most powerful businessmen. At his peak, he controlled roughly a third of Britain's railway network, sat in Parliament and was known as the Railway King.
His companies paid generous dividends, which supported their share prices and made it easier to raise more capital. In 1849, investors discovered that some of those dividends had come from money paid in by new shareholders rather than operating profits. The railways were real and the track was being built. The payouts were also helping to keep the financing machine alive.
Railway Mania did not end because Britain stopped believing in trains. The Bank of England raised rates in 1847 during a wider financial crisis. Many investors had paid only part of the cost of their shares and still faced future capital calls. When liquidity tightened, they could not pay. Forced selling followed. Hudson was exposed and disgraced, but much of the railway network he helped build powered the British economy for generations.
Lucent, Winstar, and the money that went in a circle
Move forward 150 years and the telecom boom shows the same mechanism in a form that looks much closer to AI today.
Lucent sold telecom equipment to new carriers that wanted to challenge the incumbents. Many had ambition but very little cash. Lucent solved that problem by lending them the money to buy Lucent equipment.
Vendor financing can support real customers, but it also blurs the line between demand and credit creation. Lucent shipped equipment, booked the sale and kept the customer's default risk on its own balance sheet. Part of the reported demand existed because the supplier financed it.
Winstar Communications was one of those customers. Lucent extended the credit, Winstar bought the equipment and Winstar filed for bankruptcy in April 2001. Lucent was left with losses on the receivables and loans. Its shares had peaked near $84. They eventually traded below a dollar.
Global Crossing and Qwest pushed it further by swapping fibre capacity. Each bought capacity from the other and both booked revenue, even though very little new customer demand had been created. At the same time, Global Crossing founder Gary Winnick bought a $94 million Bel Air estate and sold hundreds of millions of dollars of stock. The company went from roughly $47 billion of market value to bankruptcy in about two years. Winnick was not charged and kept the house.
The bandwidth demand was real. The financing and accounting around it were not built to last.
Cisco: right company, wrong price
Cisco is the cleaner warning because it was a real company with real customers and a real product. It built the routers that ran the internet and kept growing after the crash.
The problem was the price. At the March 2000 peak, Cisco traded at roughly 200 times forward earnings. The shares fell about 90% and took 24 years to regain that high. The business survived. The valuation did not. Investors who bought the right company at the wrong price waited a generation to get their money back in nominal terms.
Today's multiples are nowhere near Cisco's dot com peak, but they are not cheap.
So what actually causes the burst?
Across these episodes, three things usually happen in the same order.
First comes leverage with a duration mismatch. Long lived assets are funded with short term or callable capital. It happened with railway instalments, telecom networks and mortgage securities financed overnight in repo. The asset can be sound while the funding is fragile.
Second comes credit withdrawal, not an immediate collapse in demand. Customers do not suddenly stop wanting the product. Lenders stop wanting the exposure. Rates rise, a sponsor leaves, a rating falls or collateral drops. The marginal borrower can no longer refinance.
Third, vendor financing reveals how much credit was hidden inside the revenue. When suppliers finance customers, they create a sale and a counterparty risk at the same time. Demand looks strong until the buyer cannot pay. Then part of the revenue turns out to be a loan and part of the loan becomes a loss.
This is not just historical pattern matching. A July 2026 Bank for International Settlements paper compared the AI investment race with earlier technology booms. On its measures, the current build has scaled faster than canal mania, railway mania, electrification and the dot com boom at the same stage. The comparison is not perfect, but the pace and the financial links are exceptional.
On the BIS measures, AI is already at the extreme end of past investment booms.
The view today
Now apply that framework to AI. The market quickly separates into companies with very different funding risks.
Vendor financing is no longer a side issue. The BIS mapped equity stakes, purchase commitments and financing links across eleven companies. Nvidia sits near the centre. It invests in AI labs and infrastructure providers that buy its chips, supports leases and helps finance purchases. In August, Nvidia partnered with six major financial institutions on platforms designed to mobilise more than $500 billion of third party capital while keeping an option to backstop part of the financing. The loop is easy to see: Nvidia to OpenAI, OpenAI to Oracle cloud capacity, Oracle back to Nvidia hardware, with CoreWeave linking several parts of the chain. Nvidia is an investor in CoreWeave and OpenAI is a customer. The same dollar of capital can show up as equity value, revenue or backlog at several points in the network.
The BIS mapped eleven linked companies. The same capital can reappear as revenue or backlog across the network.
The concentration is what worries me. Oracle's remaining performance obligations are now roughly one and a half times its market value, and S&P estimates that more than half relate to OpenAI. Much of that backlog will only become cash over several years, while Oracle has to fund construction now. S&P cut Oracle to BBB minus, one notch above speculative grade, and the cost of credit protection moved to levels not seen in years. Blue Owl also walked away from a proposed data centre financing, reportedly because of the scale and pace of spending. Equity has already been repriced. The real question is credit quality.
Backlog is not cash. The timing and the quality of the customer matter.
The customer carrying much of the structure is also losing money. OpenAI's estimated revenue growth has slowed while its losses and infrastructure commitments continue to rise. Publicly reported commitments to multiple vendors run through 2035 and exceed $1 trillion in total, although they are not all directly comparable or fully binding. The structure only works if OpenAI can keep raising very large amounts of capital and turn contracted capacity into profitable demand.
A large part of the network depends on OpenAI keeping access to capital.
Where the analogy breaks
This is where I would stop pushing the historical analogy too far. Jensen Huang makes a fair point. Earlier manias often built capacity before demand was proven. AI infrastructure is already serving real demand, and that demand is still growing. We should not dismiss that simply because it comes from a supplier with an obvious interest in the outcome.
In 1999, many telecom buyers had little revenue. The carriers buying Lucent equipment were often funded by business plans. A meaningful share of the apparent demand came from other network builders rather than end customers.
Today is different. The largest buyers are some of the most profitable companies ever created and can fund substantial spending from operating cash flow. Microsoft and Google are reporting recognised cloud revenue from paying customers, not just bookings.
Independent pricing tells the same story. GPU rental rates fell when the market expected abundant supply, but they have risen this year across several chip generations, especially the newest hardware. High bandwidth memory remains constrained and power connections are delaying equipment that has already been purchased. Shortages and higher spot prices do not prove that every project will earn its cost of capital. They do show that current computer demand is real.
Third party spot pricing is useful because no company earnings call controls it.
The part that should worry you instead
The risk has moved. I am less worried about fake demand than financing that is getting bigger, longer dated and harder to see.
The five largest hyperscalers issued more corporate bonds last year than in the previous five years combined, and they are on pace to exceed that total again. Goldman estimates that capex is approaching 100% of operating cash flow. Morgan Stanley sees a funding gap through 2028 between planned investment and internally generated cash.
More of the build is moving from corporate cash flow into the credit markets.
The hyperscalers are not running out of money. They are using credit and structures outside the main balance sheet to preserve buybacks and strategic flexibility. That moves part of the risk to lenders and infrastructure vehicles. Moody's estimates that exposure outside the main balance sheet across the six companies most exposed to AI exceeds $1 trillion. Even Alphabet, one of the stronger self funded names, reported its first negative quarterly free cash flow since listing. Self funded is a spectrum, not a yes or no answer.
Credit markets noticed before the equity story changed. The cost of insuring Nvidia debt roughly doubled in two months, with similar moves at Meta and Alphabet. The absolute levels are still far from distress, but the direction matters. Lenders have started charging more for the balance sheet consequences of the build.
This is a duration and transparency problem, not a call that AI demand is fake. That is where the 2008 comparison becomes uncomfortable. Useful assets can still be financed badly. The BIS makes the network risk clear. Circular stakes and debt can transmit one company's failure across counterparties, while specialised AI hardware may be worth far less than book value in a forced sale.
Credit is now pricing each balance sheet on its own merits. Even the strongest names have moved.
The rehearsal
July gave us a small preview of how this can play out.
Leopold Aschenbrenner, a former OpenAI researcher still in his twenties, built Situational Awareness around an early and largely correct call on the AI investment boom. Contemporary reports put the fund's net asset value at about $45 billion after a 439% gain through June. He got the direction right and was early enough to make an extraordinary return.
The portfolio was also reportedly levered by as much as four times. When AI linked equities sold off and software shorts rallied, margin calls forced the fund to liquidate its public book in a block trade. It lost 67% in July. The thesis survived. The financing did not.
Citadel bought much of the forced portfolio, including CoreWeave, IREN and SanDisk, at distressed seller prices. It reportedly avoided Oracle's bonds. That distinction matters. Good assets sold by a leveraged owner can be attractive even when the debt financing the wider loop is not. It also fits Ken Griffin's habit of buying rare assets at auction, from a first edition US Constitution to a stegosaurus called Apex. He seems to like valuable things when the seller has limited options, whether the seller is at Sotheby's or on the other side of a margin call.
That is the whole lesson in one trade. The thesis was right, leverage forced the sale and the buyer who separated asset quality from funding risk captured the opportunity.
Where this ends
If history repeats, I think the break starts in credit, not earnings. AI does not need to stop working. The trigger could be an Oracle downgrade, a neocloud that cannot refinance or a private credit vehicle marking down data centre paper. Forced selling starts in the leveraged layer while cash rich hyperscalers remain operationally sound.
My base case is a slow absorption rather than one dramatic crash. Demand keeps growing. Commitments are renegotiated or extended. Some leveraged operators fail, recapitalise or get acquired. The gap between self funded and debt funded businesses widens. We can already see that separation in the relative performance of neoclouds and the largest platforms.
The separation has started. It is following the balance sheet, not the theme.
The more serious downside case is narrower. OpenAI's estimated revenue fails to accelerate again, an IPO is delayed or priced below what the financing structure needs, and counterparties start revising commitments or collateral values. Much of the leveraged layer sits downstream of that single funding question.
My probability weighting across four realistic outcomes.
There is one quieter signal worth watching. Berkshire Hathaway is holding record cash after several years as a net seller of equities. Buffett may not be forecasting a crash. He may simply be refusing to pay current prices. He was criticised for the same discipline in 1999. The lesson is not to copy his cash allocation. It is to separate the price you pay from your conviction in the technology.
What we take from it
Own the build where the balance sheet can carry it. Microsoft and Amazon can fund substantial investment from cash they already earn, which makes them less exposed to the circular financing problem.
Avoid the centre of the financing loop. I would not own Oracle investment grade paper or CoreWeave high yield debt at the current risk and reward. I would also be very careful wherever repayment ultimately depends on one AI lab continuing to raise capital. Credit is repricing that risk in real time. There is no reward for being early to weak paper.
Prefer the enabling infrastructure: power, transmission, cooling and memory. Those revenues are less dependent on which lab or model wins, and the evidence of real shortages is clearest in this part of the stack. Valuation and contract quality still matter.
Treat leases and guarantees outside the main balance sheet as debt when sizing exposure. The economics matter even if the accounting presentation is delayed or fragmented.
The AI story is true. The railway and internet stories were true as well. The investable question is not whether technology matters. It is who owns the paper when credit tightens. This time, the network is visible enough for us to separate durable assets from fragile financing before the break.
*Disclaimer: This report is published by Cypher Capital (BVI) Limited, a Business Company incorporated in the British Virgin Islands. Cypher Capital (BVI) Limited is not licensed or regulated by the Central Bank of the UAE, the Securities and Commodities Authority of the UAE, or the Virtual Assets Regulatory Authority of Dubai.
This report is provided for informational and educational purposes only and does not constitute investment advice, a recommendation to buy or sell any asset, or an offer or solicitation to invest in any fund, product, or strategy.
This report contains forward-looking statements and third-party price forecasts subject to significant uncertainty. Third-party forecasts cited reflect the views of those institutions, not of Cypher Capital. Cypher Capital, its affiliates, and employees may hold positions in the assets discussed herein.
No representation or warranty is made as to the accuracy or completeness of the information contained herein. Recipients should conduct their own independent analysis and consult qualified advisors before making investment decisions.
Copyright 2026 Cypher Capital (BVI) Limited. All rights reserved.*