This year alone, the total CAPEX on AI infrastructure from Big Tech companies is expected to reach 330 billion $, with Microsoft accounting for more than one-third of it, as per the latest results and projections. Where is all this money going to be spent? To buy chips and build new datacenters to serve the “overwhelming demand” for AI products and services. Now, it is important to understand the definition of CAPEX very well before moving forward:
CAPEX (Capital Expenditure) refers to funds used by a company to acquire, upgrade, or extend the useful life of physical assets that provide long-term benefits, typically beyond the current fiscal year. To qualify as CAPEX, an expense must be incurred to acquire or improve a tangible or intangible asset that will be used in the business for more than one accounting period; this means the expenditure creates or enhances a future economic benefit, such as purchasing new machinery, constructing a building, or significantly upgrading software systems, and these costs are capitalized on the balance sheet and depreciated or amortized over their useful life.
Within CAPEX there are, generally speaking, two categories: those needed to maintain, upgrade, and support revenue growth of existing businesses, and those needed to enter a new market and access new revenue streams. Now, let’s take a look at the latest Microsoft 10-K. While on one side the company continues claiming it’s investing most of its CAPEX to scale AI, the numbers do effectively provide a different story:
- “Productivity and Business processes” revenues grew 13% YoY, or by 14 billion $
- “Intelligent Cloud” revenues grew 21% YoY, or by 18.8 billion $
- Total CAPEX grew 45% YoY, or by 20 billion $
- Depreciations and Amortizations (D&A) grew 53% YoY to 34.15 billion $
It doesn’t take a degree in financial valuations to understand that if the revenue growth is much lower than the growth in CAPEX and D&As, both in absolute and in relative terms, then the company is losing money to grow its revenues, not the opposite, right?
A similar pattern can be observed in Google (post) and META latest earnings. Hold on, then, how come all these companies can grow their margins and net income? Because even if the equipment and investments accounted as CAPEX should be amortized in 2-3 years for chips or 5-6 years for data center infrastructure, as pointed out by Jim Chanos today, the D&A schedule applied is stretched way beyond that.
Here is another way to interpret this. Since AI revenues aren’t coming through as fast as all these companies expected, these are now stretching the costs well into the future, waiting (in some cases, praying) for those revenues to materialize.
I have a question now: will these Big Tech companies be able to keep growing revenues without increasing their CAPEX every year at a much higher growth rate? So far, the numbers say no, and it should not be a surprise at all because if the D&A schedule applied was the much shorter one than what’s being used by these companies, the impact on net earnings would have already been evident, and as Jim Chanos correctly stated, these companies are overstating their present earnings. As a consequence of overstating earnings today, it means that future earnings will be significantly lower than what people realize because these companies will be forced to continue accounting for D&A costs when the related items they refer to are already obsolete and no longer in use, while the new replacement costs to upgrade obsolete equipment and infrastructure will start piling up.
If this wasn’t enough by itself, we even have cases like Microsoft where, as I described last year in “MICROSOFT REVENUES “ROUND TRIPPING” PONZI SCHEME IS NOW TOO BIG TO HIDE – THE TRUTH FROM THE CASH FLOWS” and “THE SMOKING GUN THAT PROVES HOW OPENAI IS MICROSOFT’S REVENUES LAUNDROMAT”, the company is proactively round tripping CAPEX expenses into revenues using sophisticated accounting gimmicks. Things on this front might have gone a bit out of hand at Microsoft, considering that, as you can see here, the company reported a ~7 billion $ increase in “Unearned Revenues” to ~67 billion $ (an amount completely unusual and above and beyond comparable companies), but in its cash flow statement the company reported only a ~5.4 billion $ increase. What is a 1.6 billion $ discrepancy nowadays? Peanuts, right?
FY 2025 Unearned Revenues
FY 2025 Cash Flow Statement
The staggering amount of AI infrastructure investment by Big Tech companies reveals striking parallels with the telecom bubble of 1999-2000, particularly resembling the trajectories of WorldCom and Global Crossing. These parallels manifest in three critical dimensions: unsustainable spending patterns, accounting manipulations, and questionable demand assumptions—all hallmarks of previous financial disasters now reappearing in modern guise.
Today’s hyperscalers exhibit the same “build it and they will come” mentality that doomed telecom infrastructure companies. Microsoft’s 45% YoY CAPEX growth dwarfs its 21% cloud revenue growth, while Meta plans to spend 72 billion $ on AI infrastructure in 2025 —equivalent to one-third of its 2024 revenue. This mirrors Global Crossing’s aggressive 22.4 billion $ network build-out despite controlling just 15% of transatlantic capacity by 2002. Telecom companies collectively spent 121 billion $ in 2000 (approximately 213 billion $ today) on fiber networks based on projections of internet traffic doubling every 100 days—a statistic later proven to be exaggerated. Similarly, Big Tech justifies current spending with vague promises of “overwhelming AI demand” despite limited evidence of near-term monetization at scale. The capital intensity shift is profound: hyperscalers now spend 20%+ of sales on CAPEX—double their decade-ago levels—transforming from high-margin software businesses into quasi-utilities.
The creative accounting practices bear a disturbing resemblance to WorldCom’s fraud. While not identical in method, the effect is similar: artificially boosting current earnings while deferring costs. WorldCom famously capitalized 3.8 billion $ in ordinary line costs as assets to inflate profits, while today’s tech giants stretch depreciation schedules for AI infrastructure far beyond realistic useful lives. As noted before, chips that should be amortized in 2-3 years and data centers in 5-6 years are being depreciated over extended periods—a tactic that overstates current earnings by billions. More alarmingly, Microsoft’s 1.6 billion $ discrepancy between unearned revenue recognition (7 billion $ increase) and actual cash flow (5.4 billion $ increase) suggests aggressive revenue accounting reminiscent of the “prepaid capacity” entries WorldCom used to hide expenses. These practices create a temporal disconnect: inflated present earnings will collide with future D&A costs from obsolete infrastructure, much like WorldCom’s house of cards collapsed when capitalization couldn’t offset real expenses anymore.
The specter of massive oversupply looms as it did for Telecom companies in the 2000s. Telecom companies laid enough transatlantic fiber by 2000 to serve projected demand for decades, resulting in bankruptcies when utilization rates plummeted below 30%. Today, similar dynamics emerge: J.P. Morgan estimates data center capacity will grow 58% annually through 2026, while AI traffic grows at 26%—a mismatch threatening catastrophic overcapacity. SoftBank’s Masayoshi Son exemplified bubble-era thinking with his call for 9 trillion $ in AI infrastructure to support a hypothetical artificial superintelligence by 2035. Meanwhile, open-source AI advancements like DeepSeek, achieving GPT-4 performance with much fewer resources, undermine the narrative that computing power alone guarantees dominance, as I described in my article “THE REAL ERA OF AI BEGINS, THE ONE OF THE AI CHARLATANS ENDS”.
Unlike the telecom bubble, Big Tech funds investments through cash reserves rather than debt, and possesses established revenue streams. However, these advantages may be illusory: free cash flow at hyperscalers fell 15-20% in late 2024 despite revenue growth, with Morgan Stanley projecting further 15-20% declines in 2025. The path forward is fraught: if AI monetization lags expectations, today’s stretched depreciation schedules will snap, forcing massive write-downs. Investors should heed history—when infrastructure booms outpace real demand, the aftermath favors nimble, asset-light innovators (like Netflix post-telecom bust) and commodity suppliers (power/equipment vendors), not the infrastructure builders themselves. As in 2000, the greatest beneficiaries may emerge outside the frenzy—those who leverage the infrastructure without bearing its costs.
JustDario on X | JustDario on Instagram | JustDario on YouTube
SYNNAX SUMMER PROMO
Get exclusive intelligence on stocks you can’t find anywhere else – insights that could give you a serious edge. Claim your free Synnax trial month now: https://synnax.app/PROMO/JUSTDARIO
If you love what we built at Synnax (and I think you will), this advanced market intelligence will cost you less than a cup of coffee a month for the next 12 months. After the 30th of September annual subscription price will be 154.99$ – redeem your free access now and give it a try.