AI can win. Your portfolio can still lose.
Should you sell every semiconductor stock? My answer: keep selective exposure, hold substantial cash, and size positions around what happens if you are wrong.
A ten-year technology thesis does not justify unlimited risk at today's price.
Here is the framework.
1. STRONG AI REVENUE DOES NOT UNDERWRITE EVERY AI PROJECT
Imagine two futures:
A) Model revenue grows, customers generate cash, and infrastructure earns an adequate return.
B) Model revenue grows, but construction slows and expensive infrastructure projects restructure their debt.
An AI company reporting higher annualized revenue is compatible with BOTH.
That is the Bayesian point: evidence matters to the extent that it helps distinguish competing outcomes.
The evidence I want to see is customers funding purchases from operating cash, projects servicing debt after essential equipment replacement, and less dependence on the next financing round.
A cloud provider announces capex. A chipmaker announces orders. An equipment supplier announces expansion.
Those may be three observations of the same budget. Counting them as three independent confirmations creates false confidence.
2. POSITION SIZE STARTS WITH THE LOSS YOU CAN AFFORD
Take $TSM.
At the October 30, 2026 close, its ADR traded at $406.13. The last four reported quarters totaled roughly $15.54 of earnings per ADR, or about 26.1x reported trailing earnings.
Now stress it:
• Annual earnings fall 20% from that trailing figure.
• The market pays 20x earnings.
• Implied price: roughly $249.
• Price decline: roughly 39%, excluding dividends.
That is a stress scenario, not a price target.
No technological collapse is required. Earnings and the multiple simply decline together.
Suppose you allow semiconductor exposure to cost your portfolio 12% in a scenario where the sector falls 40%.
12% / 40% = 30% semiconductor exposure.
In a simplified stocks-and-cash portfolio, that leaves 70% cash.
Stress a 50% decline instead, and the same loss budget supports only 24% exposure.
These are illustrative allocations, not an optimal portfolio for everyone or a guarantee that losses stop there.
For a single stock, a 60% downside scenario and a 2% portfolio loss budget imply a position of about 3.3%.
Put 30% of your portfolio in that same stock, and the scenario costs you 18%.
Conviction does not change the arithmetic.
Nor does owning five tickers diversify you if all five depend on the same hyperscaler spending cycle.
3. WHICH SEMICONDUCTOR RISKS ARE WORTH KEEPING?
My research priority is businesses that can fund themselves through a downturn and retain pricing power.
$TSM: a manufacturing platform serving competing chip designs. Changes in the winner among GPU vendors and custom chips need not destroy the value of advanced manufacturing. Geopolitics, overseas costs and the capex cycle still belong in the risk budget.
$NVDA: exceptional economics from chips, systems and software. But fabless does not mean free of customer spending risk, receivables risk or supply commitments. Outsourcing fabrication does not outsource the business cycle.
$MU: distinguish financial strength from earnings durability. Strong cash generation can support debt service while memory prices, new supply and customer inventory threaten profits. Long-term contracts do not automatically lock in peak margins.
Using October 30 prices and reported trailing earnings, TSMC and Nvidia were around 26x earnings; Micron was around 12x.
Those are historical accounting multiples, not normalized earnings valuations.
If Micron's annual earnings were to halve while its share price stayed unchanged, 12x would become 24x. That is sensitivity analysis, not a forecast.
The question is how long the earnings you are buying will last.
4. SOFTWARE IS NOT AN AUTOMATIC SAFE HAVEN
Moving from semiconductors into $GOOGL or $FIG can exchange one risk for another.
Google's search business is still growing. The investment question is where high-intent users will make decisions, where advertisers will acquire customers, and how much cash remains after AI infrastructure spending.
Figma is growing too. AI can expand its user base and usage revenue while changing the economics of paid seats. Growth alone does not settle what happens to margins and cash generation per share.
At a price that compensates for those uncertainties, a controlled position can make sense. Without that compensation, zero is a valid allocation.
What I reject is allowing one unproven business-model assumption to determine the fate of the entire portfolio.
Even the Berkshire example requires the right denominator.
At June 30, 2026, its reported Alphabet holdings were roughly $37.8 billion:
• About 12.6% of its 13F portfolio.
• About 3.0% of group assets.
• About 5.0% of equity attributable to Berkshire shareholders.
That is a major equity position, but a different exposure at the conglomerate level. Group assets also include assets financed by liabilities.
Copying the ticker without copying the balance sheet produces a different trade.
5. CASH NEEDS A REDEPLOYMENT RULE
Cash lets you buy without first selling another asset into the same drawdown.
A data center can bankrupt its expensive first capital structure and become profitable for a buyer who acquires it at a lower cost. An industry does not have to repay every early investor at the original price to succeed.
Waiting has a cost. If stocks are 30% of your portfolio and rise 20%, the contribution to the whole portfolio is only 6%, before cash interest.
A defensive strategy must accept that trade-off.
But "wait until everything is safe" is not a strategy either.
I would redeploy when:
• Price falls enough to improve prospective returns, while the business and customer payments remain intact.
• Operating evidence improves: less refinancing dependence, cash left after necessary equipment replacement, and profits that are not being swallowed by receivables and inventory.
• The resulting position still fits the portfolio's loss budget.
I would reduce exposure when financing or payment quality deteriorates without enough compensation in the price—or when a rally pushes the position beyond its risk budget.
You do not need to predict the date of a crash to reduce excessive exposure.
Before buying, answer three questions:
What earnings does today's price require?
What does the bad scenario do to my portfolio?
How much deployable cash remains afterward?
Being right about the industry is only the beginning of being right about the investment.
Original analysis by 硅facts. Personal views, not investment advice.
AI Revenue Can Grow While AI Debt Blows Up
AI revenue doesn't have to collapse for AI debt to blow up.
My base case: leading model companies keep growing revenue. Some of the infrastructure built around them still defaults.
Global Crossing filed for bankruptcy in 2002. It had issued $3.8 billion of senior notes originally due in 2006–2009. Long-dated debt did not buy it enough time. The network kept operating; the old shareholders were wiped out in the restructuring.
Investors were right about the internet. That did not make every price, project or balance sheet right.
AI runs on three clocks: models improve quickly, organizations change slowly, and debt requires payment on schedule.
A company can buy AI today to defend its market share. The model provider gets paid immediately. Turning that purchase into more output may require years of changes to workflows, permissions, accountability and physical equipment.
Some AI revenue also comes from existing software, consulting and outsourcing budgets. That is real revenue. It just isn't all new economic activity.
This is how model revenue can surge while broad productivity gains arrive gradually over a decade. The construction boom already adds economic activity; whether customers ultimately earn enough to support those assets is a separate question.
Now consider the part that matters for investors: revenue can keep rising while new equipment orders fall.
A simple hypothetical: assume revenue and required server capacity grow in proportion, with no equipment retirements.
Required servers: 100 → 150 → 180.
New servers purchased: 50 → 30.
Revenue grows 20%. New equipment orders fall 40%.
The business is still growing. Its expansion orders are shrinking. A project financed for continued acceleration can get into trouble without a collapse in AI demand.
Efficiency can sharpen that split. Cheaper inference may improve a model company's margins while putting pressure on rental rates for older hardware. The customer's savings can become the asset owner's lost income.
Contracts decide where the pain lands. Minimum purchase commitments keep payment obligations with the buyer. Flexible orders leave more risk with the infrastructure provider. “AI revenue is growing” tells you very little about which borrower can cover its debt.
My call for the next three to five years: continued adoption, growing revenue at leading labs, and restructurings among projects financed on aggressive assumptions about growth, rental prices and equipment resale values.
The strongest operators can use improving cash flow to fund themselves—or buy distressed assets cheaply. A wider credit crisis requires those losses to spread into a broader lending pullback.
Watch who keeps cash after operating costs, necessary equipment replacement and debt service. A token chart cannot answer that question.
The next owner may run the same data center profitably after buying it for a fraction of the original cost.
AI doesn't owe the first owner their money back.
Why Anthropic Could Become the Biggest Stock-Market Bubble in History
At a potential $2 trillion valuation, investors would be paying today for many generations of model leadership that Anthropic has not yet won.
That is the bet: Claude keeps delivering enough economic advantage to justify premium pricing, perhaps even reaching AGI first. A great product today cannot secure that outcome.
Here is the risk buried beneath the growth story.
Open models do not need to beat Claude at everything.
They only need to become good enough for a large share of paid workloads. Enterprises can keep Claude for their hardest problems while moving routine tasks to cheaper alternatives.
Anthropic could retain its best customers and still lose pricing power. Even customers who stay gain leverage to demand discounts.
AI usage can boom while the economics of selling intelligence deteriorate.
Model leadership can change faster than infrastructure contracts.
Anthropic’s S-1 describes minimum spending commitments of roughly $114.5 billion to Google through 2033 and $110 billion to Amazon through 2036, subject to contractual exceptions.
Around 80% of first-half revenue was usage-based. Some customers have minimum commitments, but customer spending and upstream obligations are not perfectly matched.
If prices fall or workloads migrate, previously scarce capacity can become an expensive obligation. Meanwhile, cutting research to conserve cash risks falling further behind.
Financing can keep the company alive while destroying shareholder returns.
The Broadcom-related financing facility illustrates the mechanism. Up to $42 billion can offset specified lease payments through notes; it is not unrestricted cash.
Normal terms include 9% interest and conversion at 90% of the reference share price. Notes issued when specified revenue conditions are unmet can carry 13% interest and convert at 85%.
Those conditions have not been shown to be triggered. But the structure matters: weaker business performance can mean more expensive financing and greater potential dilution.
A $2 trillion entry price leaves little room for ordinary success.
As an illustration, even a generous 40× multiple would require $50 billion of annual cash available to common shareholders to support that valuation. At an assumed 25% distributable cash margin, that means $200 billion in annual revenue.
These are scenario assumptions, not forecasts. They also ignore the cost of waiting and future dilution.
Anthropic could become a huge, profitable company and still be a terrible investment at that price.
The strongest counterargument is real: Anthropic has substantial cash reserves, enterprise relationships, improving operations and some cancellable compute contracts. One disappointing model release would not wipe it out.
The extreme downside requires persistent competitive failure, cash depletion and failed refinancing or restructuring. Common equity can then be wiped out even if the technology and business survive.
AI does not have to fail for ANT shareholders to lose.
The biggest bubble could form around a genuinely transformative company, because investors confuse the importance of its technology with the value of its stock.
At $2 trillion, you are buying the expectation that Anthropic keeps winning. The contracts may last longer than its lead.
Valuation caveat: $2 trillion is a reported potential valuation, not a confirmed IPO price. “Biggest bubble” is a risk thesis, not a forecast or a calculated probability.
Why US Stocks Could Go Nowhere for Years
Imagine sitting at the White House economic policy table. Three problems are in front of you: technology companies want to keep building data centers, voters want a lower cost of living, and the Treasury needs to keep selling debt.
You want AI investment to continue. It supports growth, jobs, and technological leadership. You also want lower interest rates, because households, businesses, and the government all find borrowing expensive. Yet the faster investment expands, the more pressure it puts on power, equipment, and capital. Cut rates too aggressively, and bond investors may start worrying about inflation again.
Move to the Federal Reserve, and the trade-offs remain. You can believe AI will raise productivity without treating tomorrow's promised cost savings as an answer to today's inflation. You need a functioning financial system, but that does not require protecting every elevated valuation.
Policymakers in both buildings need time. That is why I think US stocks could face years of disappointing returns: AI's success has yet to deliver broad enough economic gains to validate the expectations embedded in asset prices, while evidence of failure remains too limited to overturn the entire investment cycle.
Rallies run into inflation and debt constraints. Selloffs run into real demand, corporate cash flow, and potential policy support. Repeated repricing becomes more plausible than a durable move in one direction.
Washington wants the buildout to continue
This tension is already visible in policy.
In March 2026, the White House secured commitments from major technology companies to bear the costs of additional power generation and related grid upgrades for data centers, with the aim of protecting households from higher electricity bills. The political objective was straightforward: keep building AI infrastructure while containing the cost to voters.
But assigning the bill does not instantly produce another transformer, skilled electrician, or power plant. If companies absorb more of the cost, investment returns come under pressure. If households absorb it, purchasing power and political support suffer. Someone has to pay for scarce resources.
In September, the Fed acknowledged strong productivity growth and robust capital investment while raising its policy rate to 3.75%–4.00%. September CPI, released in October, showed prices up 3.7% from a year earlier. That offered little comfort to investors counting on rapid easing.
The sequence matters. AI first buys equipment, builds facilities, and competes for energy. Broader efficiency gains arrive as businesses redesign how they operate. The spending increases demand before the full productivity dividend expands supply.
The White House wants to make that future dividend possible. The Fed has to manage the prices being paid in the meantime. The same AI development can therefore improve a technology company's revenue outlook and delay the rate cuts investors want.
For assets already priced for years of growth, good news does not automatically justify a higher valuation.
Earnings need to catch up with promises
From the White House's perspective, AI's most valuable outcome is an economy that produces more saleable goods and services with the same resources. That would support profits, household incomes, and the tax base. Spending more on data centers is only an input into that process.
Productivity is already improving. US nonfarm business labor productivity rose 2.2% year over year in the second quarter. The next question is whether those gains can spread and persist, improving businesses beyond the AI sector quickly enough to justify the investment and valuations already committed.
Faster coding is a beginning. Changes to procurement, logistics, healthcare, manufacturing, sales, and management are what gradually reshape costs across the economy. Models can improve every month. Organizations, infrastructure, and working practices rarely move at the same speed.
Federal borrowing makes the waiting expensive. In its early October estimate, the Congressional Budget Office put the fiscal 2026 deficit at almost $2 trillion. The Treasury must finance new borrowing while refinancing maturing debt. The White House would welcome cheaper funding, but a lower Fed policy rate does not guarantee lower long-term Treasury yields.
Bond investors still price inflation, the future path of rates, and the risks of holding long-duration debt. If they think easing will tolerate more inflation, long yields may refuse to follow short rates down. Larger deficits used to support demand can also raise the return investors require.
For expensive equities, that means distant profits continue to be discounted at a demanding rate.
A sustained advance therefore needs a credible path toward broader earnings growth, easing inflation pressure, and a more manageable fiscal trajectory. The federal debt does not have to disappear. Investors do need to believe that income and productive capacity can increasingly support the obligations being accumulated.
Until that path becomes clearer, rallies risk spending the productivity dividend before it arrives. Rising asset prices can then reinforce wealth effects and investment, supporting demand and bringing the constraints back sooner.
An AI bust needs its own dark-fiber reckoning
Policymakers also need more than a large investment number to conclude that the entire buildout is a mistake.
The telecom industry showed what a genuine breakdown looked like around 2000. Companies laid vast amounts of long-haul fiber. Paying demand failed to grow as fast as builders had expected, while construction and technological advances rapidly expanded transmission capacity.
The internet kept growing. Customers gained cheaper communications. Some of the companies financing the infrastructure saw their capital returns destroyed.
Dark fiber became a symbol of that overbuild. For lenders and shareholders, the central issue was whether assets could earn enough, soon enough, to repay the capital committed to them. Eventual usefulness could not meet an approaching debt payment.
An AI version would likewise show up in investment returns: additional capacity persistently selling at prices too low to cover its costs, customers cutting or renegotiating purchases, falling equipment resale values, and increasingly difficult refinancing. As these pressures reinforce one another, isolated project failures can become an industry-wide contraction.
GPUs face an especially unforgiving timetable. New chips improve energy efficiency and cost per task. Older equipment can lose economic competitiveness while its owners wait for demand to catch up. A data center can remain busy while its shareholders fail to recover their original investment.
Markets will anticipate that deterioration. They do not need to wait for facilities to close. But a sustained, 2000-style collapse along the AI investment chain requires growing conviction that the revenue generated by additional capacity cannot justify continued expansion.
Until that conviction strengthens, the established cash flows of large technology companies, actual customer spending, and genuine technical progress will keep challenging the case for wholesale failure. Individual borrowers can get into trouble well before the industry loses its economic value.
That creates a frustrating middle ground: AI is useful enough to keep construction going, while the gains spread too slowly to justify ever-higher valuations.
Policy can buy time while shareholders pay for it
In this situation, the White House has reason to accelerate power supply, permitting, and infrastructure delivery. The Fed must adjust to inflation, employment, and financial conditions.
Their responsibilities differ. Neither is required to defend a particular stock index level. Both, however, have reason to avoid a sudden collapse in financing for viable projects or a breakdown in markets that damages the wider economy.
That makes an adjustment in stages plausible. When activity overheats, financing costs and policy constraints restrain expansion. When growth weakens and inflation eases, room for monetary easing returns. If credit markets malfunction, liquidity tools may come into play. Restoring market functioning does not require restoring yesterday's equity valuations.
Policy could therefore prevent an uncontrolled financial contraction while leaving investors with several years of poor stock returns. Washington gains an adjustment window. The Fed preserves its inflation credibility. Shareholders absorb part of the cost through lower valuation multiples and the passage of time.
Profits do not even have to stop growing.
Suppose an index begins at 30 times earnings. Earnings grow 8% a year for three years, increasing by roughly 26%. Over the same period, its valuation falls to 24 times earnings. The index price ends up almost exactly where it started. Dividends are additional, but much of the operating progress has gone toward absorbing the price investors paid at the beginning.
That is the dead-money scenario: businesses evolve, winners and losers diverge, and stocks experience powerful rallies and painful declines, yet owning assets priced for high expectations produces little reward for the volatility endured.
Every rebound looks like a new bull market. Every correction looks like the start of a crash. Both sides keep mistaking an interim development for the final verdict.
This can last for years because productivity diffusion, power infrastructure, debt refinancing, and investment returns operate on longer timelines than a Fed meeting. Quarterly data can change the market's mood without resolving any of those processes.
The clearest routes out lie in two sets of evidence. One is a broad improvement in earnings and incomes beyond the AI sector, accompanied by easing inflation and fiscal financing pressure. That would strengthen the case that growth can justify valuations.
The other is a persistent deterioration in returns on additional computing capacity, followed by weakening customer commitments and refinancing access. That would force investors to recognize that capital needs to be written down.
Until one path becomes more convincing, treating every rally as the start of a lasting bull market, or every selloff as another 2000, risks committing too early.
For the White House and the Fed, getting the economy through this transition could count as a policy success. For investors, success requires an additional condition: paying a price that leaves room for the wait.
Washington needs time. Shareholders need returns. The next few years may expose just how far apart those two objectives can be.
Original analysis by 硅facts. No reproduction without permission. The views expressed are personal and do not constitute investment advice.
The AI Revenue Bear Case Is Too Focused on Token Prices
Customers can switch models without giving up AI. And today's entry-level subscription price tells us surprisingly little about what a mature customer might eventually spend.
That is the monetization upside I think investors risk underestimating.
Consider Barclays. In an October 1 update, Anthropic described a Claude-powered platform in the bank's Global Markets business that processes roughly 120,000 emails a day, helping classify requests, add information and route them into the appropriate workflows.
Imagine switching it off tomorrow. The emails would keep arriving. Someone, or another system, would still have to do the work.
Barclays can negotiate harder or change vendors. But once AI is embedded in a productive workflow, eliminating the software bill can create a cost somewhere else.
That changes the economics of customer demand.
The cost of going back matters
Imagine a small software team that starts using AI to explain code and draft tests. Initially, canceling the subscription is easy. The team simply returns to its old process.
Over time, suppose the tools prove useful after accounting for review and rework. Developers integrate them into their repositories, learn where they need supervision and start completing more work with the same headcount. Delivery schedules adjust accordingly.
Canceling now means accepting slower releases, doing less maintenance or adding labor.
The relevant comparison has expanded from a software subscription to the cost of completing the work without it.
This is why I am more constructive on the leading AI labs' aggregate demand base than a falling token-price chart alone would suggest. Market share can move between vendors while the underlying workflow remains dependent on AI.
Prices can still fall enough to depress aggregate spending. Durable usage gives the industry a revenue foundation; each vendor must earn its share of that spending.
Monetization can improve without an across-the-board price hike
Low entry prices make experimentation easier. Customers often need to use a product before they can identify its most valuable applications.
If those applications become part of daily work, the next stage is greater willingness to pay for capacity, reliability and execution.
Consider a hypothetical customer base of 100 users paying $20 a month. Monthly revenue is $2,000.
If 10 users upgrade to a $100 tier while everyone else stays on the original plan, revenue rises to $2,800: 40% growth, with no increase in the entry price and no new customers.
That is a mix shift, not a same-product price increase. The distinction matters. Higher tiers also bring higher service costs, so the revenue uplift still needs to translate into gross profit.
Enterprise accounts offer more ways to expand: additional seats, paid usage, stronger administrative controls and service commitments. A task that can wait overnight has different economics from one blocking a product launch.
The cheapest API endpoint is a poor proxy for the revenue potential of the entire customer relationship.
Falling prices can coexist with improving gross profit
AI has a real usage cost. A subscriber who becomes more engaged can consume enough compute to overwhelm the extra revenue.
The investment case therefore requires progress on both monetization and delivery efficiency.
Caching repeated context, routing suitable tasks to smaller models and batching requests can reduce the resources needed to deliver a given service. Better reliability can also reduce retries and human intervention. These gains must hold at comparable task quality.
Here is an illustrative example, not an estimate of any company's margins:
A service generates $20 of revenue and costs $12 to deliver. Gross profit is $8, with a 40% gross margin.
If delivery cost falls to $8, the company could cut its price to $18 and still earn $10 of gross profit. The customer pays 10% less; the provider earns 25% more gross profit per customer.
The key relationship is the rate of cost improvement relative to price erosion.
Combining that with premium-tier adoption and account expansion creates a credible path to higher gross profit even while basic intelligence keeps getting cheaper.
Underwrite the customer economics, then the capital commitments
The strongest potential businesses combine capable models with products embedded in real work. Replacing a simple API call may be easy. Replacing a system connected to permissions, tools and operating procedures can require meaningful validation and migration effort.
That can support retention and pricing power, provided the product keeps delivering value.
For investors, the useful questions are concrete: Do existing customers spend more over time? Does gross profit from those customers grow after the full cost of serving them? How much recurring R&D is required to keep the product competitive? How far ahead of revenue must the company commit cash to capacity?
An annualized revenue run rate is not contracted SaaS ARR. Gross profit is not free cash flow. And a useful product does not justify an unlimited valuation.
But those distinctions should sharpen the investment analysis, rather than end it at the latest API price cut.
My base case is that leading labs have more room to monetize established workflows than today's entry prices imply. Premium adoption, enterprise expansion and lower delivery costs can reinforce one another.
The thesis fails if a mature customer base still cannot generate enough gross profit to fund the ongoing investment required to stay competitive. In that case, customers are capturing the value while the labs struggle to retain it.
Back at Barclays, the next day's emails still need processing. The commercial contest is over who can do that work reliably, at an acceptable price, while keeping enough profit to improve the service.
That is the business I want to underwrite. Today's token price is one input. The lifetime economics of the customer relationship determine what the business is worth.
Original analysis by 硅facts. No republication without permission. Personal views only; not investment advice.
最近在硅facts微信公众号和抖音,系统分析了��什么NAND会更紧缺,DRAM二阶导已经转负,既是成本问题,也是Infra技术问题。欢迎订阅,微信是单篇付费,抖音是按月付费+会员社群。知识变现,不喜勿喷。Just published a deep dive on the SiliconFacts (硅facts) WeChat & Douyin:We analyze the systemic reasons behind the upcoming NAND shortage and why DRAM's momentum (second derivative) is turning negative. It's a complex mix of unit cost economics and underlying Infra tech constraints. 💡#Semiconductors #NAND #DRAM #AIInfrastructure #TechTrends
$SNDK $MU $SKHY
A lot of you are confused between HBF and HBM. Worse, some of you pretend to know and write posts that mislead everyone else.
They will never replace each other. Here's why:
As a memory expert, I could write an IEEE style journal on this right here. But it would just get stolen by the big accounts, so I won't.
Keeping it simple.
The easiest way to think about HBF and HBM is to think about NAND and DRAM relationship today. It's a symbiosis. They can't replace each other. Not with better controllers. Not with smarter firmware. Not ever.
The reason is PHYSICS.
NAND has density but wears out. It degrades over time.
DRAM is low latency but volatile. Data goes puff when power goes out.
NAND memory is called NAND because it actually uses NAND gates and they are connected serially, so it does well on serial read (big file reads/writes in which bits are done sequentially cell by cell). So HBF might be suitable to store big AI Models.
DRAM cells can be addressed individually, so it does well on random access, small to medium size files spread across different physical/logical addresses (hence the name Dynamic Random Access Memory). So HBM might be more suitable for KV cache or frequently used Models.
Memory makers aren't selling you products. They're selling you one type of physics over another. You can't swap physics. You choose one physics for your application.