On 22 October 2024, FC Barcelona beat Bayern Munich 4-1 at home. Raphinha scored a hat-trick, Lewandowski added another, and the Camp Nou roared itself hoarse. In the second half, however, Bayern had three clear one-on-ones. All three were missed. Had Harry Kane been slightly more clinical, the headline the next morning would have been “Barça humiliated in open play” instead of “Raphinha masterclass.”
That single night encapsulated everything about FC Barcelona under German head coach, Hansi Flick. They attack with a ferocity rarely seen in modern football, concede chances by the bucketload, and trust that their forwards will always score one more goal than the opposition can convert. It is beautiful, reckless, and sustainable only as long as the attack keeps delivering.
Exactly the same thing is happening right now with the $300 billion-plus annual AI infrastructure build-out by the hyperscalers and Nvidia.
The Tactical Blueprint
Flick’s system has four non-negotiable pillars.
- Relentless high press from the first minute.
- Full-backs permanently stationed in the opposition half as auxiliary wingers.
- Central midfielders timed to arrive late in the box for tap-ins.
- Defensive structure treated as an optional extra.
The result is an average of 3.8 goals scored per La Liga match this season and 1.7 conceded. Critics call it suicidal. Fans call it the most entertaining football in Europe. Both are right.
| The Translation Table | |
| FC Barcelona 2024–25 | AI Hyperscalers + Nvidia 2025 |
| Lewandowski, Yamal, Raphinha, Pedri | GPT-5, Gemini 2, Grok-3, Llama 4, Claude 3.5 |
| Full-backs bombing forward (Koundé, Balde) | Every spare dollar of cloud profit turned into capex |
| Iñigo Martínez left isolated on counters | Six-year depreciation schedules on thirty-month hardware cycles |
| Flick refusing to play three centre-backs | CFOs refusing to shorten useful-life assumptions |
| Bringing on Ferran Torres in the 80th minute | Raising capex guidance again in the next 10-Q |
| Conceding two but winning 5-2 | Reporting record EPS while quietly building a $176 billion depreciation overhang |
2025: Big Tech Is Now Spending (Almost) Everything It Earns on AI
Operating Income vs AI-Related Capex: the Hansi Flick scoreboard in dollar
Source: Company Q3 2025 earnings releases and guidance (Meta 29 Oct, Alphabet 29 Oct, Microsoft 29 Oct, Amazon 31 Oct 2025); operating income from Bloomberg/Yahoo Finance consensus as of 28 November 2025.
The Scoreboard in November 2025
The attack is still on fire.
Meta’s Q3 2025 net income rose 48 % year over year while capex guidance was lifted again to at least $70 billion for the full year. Alphabet increased its 2025 capex estimate for the third time in six months, now $91–93 billion. Microsoft and Amazon are both guiding toward $100 billion-plus. Nvidia delivered $35 billion of revenue in a single quarter and declared Blackwell sold out through 2026. Azure AI revenue grew more than 60 %, AWS growth is re-accelerating, and inference pricing has not collapsed.
Every earnings call sounds like Hansi Flick after a 4-3 win: “We know we give away chances, but we create even more. The best defence is a lethal attack.” Investors applaud and buy more stock.
The Leaky Back Line Nobody Wants to Discuss
While the forwards are stealing the show, the defence is being exposed every week.
- Depreciation overhang. Michael Burry’s November calculations (still unchallenged on the numbers) estimate $176 billion of cumulative understated depreciation across the hyperscalers from 2026 to 2028 if hardware cycles remain thirty months or shorter. The issue is simple: companies are spreading asset costs over five to six years while the hardware itself becomes economically obsolete in half that time. The gap does not vanish. It accumulates.
- Energy inefficiency. Older GPUs (A100, H100) consume two to three times more power per flop than Blackwell or Rubin. Electricity is rapidly becoming the binding constraint.
- Monetisation gap. Training spend dwarfs inference revenue by orders of magnitude. The industry is still in the “score five to cover the two we concede” phase.
- Balance-sheet red flags at Nvidia. Receivables hit $33.4 billion, inventory $19.8 billion, days sales outstanding creeping higher.
- Competitor encroachment. Google’s Ironwood TPU, Amazon Trainium2, and AMD’s MI350 series are no longer jokes. They are cheaper, more power-efficient for inference, and gaining share.
These are not theoretical risks. They are the opposition striker standing unmarked in the six-yard box every transition.
The Bad Day Scenario
Imagine one quarter in 2026 or 2027 when the attack has an off night.
The next generation of frontier models arrives and the leap is incremental, not mind-blowing. Inference pricing falls 40–60 % because supply finally outstrips training demand. A major hyperscaler is forced to impair $15–30 billion of Hopper or early Blackwell assets because the new Rubin GPUs make them economically obsolete. Capex guidance is cut 20–30 % across the board. Nvidia’s inventory and receivables suddenly look toxic.
The market reaction would be swift and brutal. The same stocks that ignored two conceded goals while celebrating five scored would suddenly see only the leaky defence. The dot-com bandwidth glut is the closest precedent: an industry built capacity for exponential demand, only to discover demand was merely linear. A 30–50 percent drawdown in the Magnificent Seven within weeks would not be surprising. Michael Burry would be carried through the streets like a returning hero.
Why They Will Never Fix the Defence
Hansi Flick will not suddenly adopt a back five and park the bus. If he did, Xavi or Luis Enrique would still be in the job.
Mark Zuckerberg, Sundar Pichai, Andy Jassy, and Satya Nadella will not unilaterally slash capex while their rivals keep spending. The first mover to play defensively loses the AI arms race and watches their market share evaporate. Game theory is merciless: everyone is incentivised to keep attacking until the scoreboard itself forces a change.
So they will raise guidance again, extend useful lives a little further if necessary, and trust that the next model is another Raphinha hat-trick.
The Incentive Trap
Nobody inside the GPU ecosystem is paid to play defence.
Nvidia’s entire valuation rests on releasing a new architecture every 18–24 months that makes the previous generation economically obsolete overnight.
Hyperscalers must keep buying because the first CEO to blink and cut capex instantly cedes frontier-model leadership and cloud market share to the others.
Frontier labs (OpenAI, Anthropic, xAI) keep demanding 10× more compute because anything less is marketed as stagnation and kills fundraising.
It is a classic self-reinforcing high press. Every participant is incentivised to push further and further up the pitch, leaving acres of space behind them. No executive wants to be the one who drops deep first and watches their rivals score in the open goal they just vacated.
The system only rewards relentless attack, even when the defence is already on its knees.
Living on the Scoreboard
As of 28 November 2025, it is the 87th minute and the scoreboard still reads 5-2. Camp Nou is rocking. The opposition has hit the post twice and missed a sitter, but nobody is talking about tactical balance.
The conceded goals only matter on the day the attack fails to outscore them.
Everyone in the stadium knows this. The players, the coach, the fans, even the opposition.
They also know that when the attackers finally have a bad day, the silence will be deafening.
Until then, enjoy the football.
