The Number That Changes Everything
If you read only one number today, make it this one: $750 billion. That's the potential computing power bill Anthropic is reportedly preparing to pay. Not over a year. Not over a decade. The scale is almost incomprehensible — it dwarfs the entire GDP of Turkey, exceeds the combined market cap of every listed company in Southeast Asia, and represents a concentration of capital expenditure that has no parallel in the history of industrial transformation.
Let that sit for a moment.
Anthropic spent nearly $9 billion on computing capacity in a single period. The jump from $9B to a $750B commitment isn't a growth curve — it's a civilization-scale gamble. And here's what no one in the breathless AI coverage cycle wants to say out loud: if even a fraction of this spending doesn't generate proportional returns, we are looking at the largest misallocation of capital since the dot-com fiber optic buildout of 1999-2001.
The historical parallel is uncomfortable. In 1999, Global Crossing laid 100,000 miles of dark fiber across the Atlantic, betting on internet traffic doubling every 100 days. It doubled — but not fast enough. The company went bankrupt. The fiber got used. The shareholders got destroyed. The infrastructure outlived the investors who funded it.
Anthropic's $750B question is the same question with a bigger multiplier.
Southeast Asia's Identity Crisis Is Now a Narrative Crisis
The Google-Temasek-Bain e-Conomy SEA report dropped its annual prophecy today: the region's digital economy will hit $1 trillion by 2035. Sounds great. Except this target slipped five years. What was supposed to happen by 2030 now has a 2035 deadline.
The old play was simple: mobile internet adoption drives e-commerce and fintech, which compounds into a trillion-dollar economy. That playbook worked — 2019 to 2024 was genuinely spectacular. But the second act requires something the region hasn't built: genuine AI infrastructure, local-language model capabilities, and the kind of deep-tech talent pipeline that Israel or Shenzhen possess.
The report notes 243 million Southeast Asians have adopted generative AI — 58% of the working population. That's an adoption rate that would make Silicon Valley envious. But adoption without monetization is just expensive curiosity. The GMV is $368 billion, up 17% — solid but not the hockey stick that justifies trillion-dollar valuations.
Meanwhile, the region's capital markets are telling a different story. DayOne is considering delaying and shrinking its US IPO. Oyo's parent Prism is cutting its fresh issue to $353 million. The exits aren't coming at the scale the fund inflows of 2021-2022 assumed they would.
The uncomfortable truth: Southeast Asia's digital economy is becoming AI-washing, not AI-powered. The $1T number persists because it's useful for everyone — governments want to attract foreign direct investment, VCs want a TAM story, and media outlets want a headline. But the actual AI value capture in SEA remains embarrassingly thin compared to the US or China.
The Quiet Chip Bifurcation No One Is Talking About
This is the story buried under the noise: DeepSeek open-sourced a tool to reduce deployment costs on both Nvidia's SM100 Blackwell GPUs AND Huawei's Ascend 950 chips.
Read that again. A Chinese AI lab is building infrastructure that works across both the American and Chinese chip stacks. This isn't neutral technology — it's a geopolitical statement disguised as an engineering achievement.
For two years, the US export control regime has operated on a simple thesis: if you can't buy Nvidia, you can't compete. The assumption was that the CUDA moat was so deep that no amount of domestic Chinese engineering could bridge it. DeepSeek's move says otherwise — or more precisely, it says that open-source software can bridge the gap between two incompatible hardware ecosystems, creating a de facto normalization of the bifurcated compute landscape.
If this works at scale, it undermines the entire export control strategy. Not by circumventing the chips themselves, but by making the software layer portable enough that the hardware choice becomes a cost optimization problem rather than a capability constraint.
The implication for US chip policy is existential. The Nvidia-Blackwell premium exists because there was no alternative path. If DeepSeek-style abstraction layers mature, that premium compresses.
Bitcoin's Thesis Is Dead. Long Live Bitcoin.
Bitcoin sits at $82,829. It's up 1.67% in 24 hours. It's tracking the S&P 500 with a 95% correlation.
If you believe Bitcoin is digital gold — an uncorrelated store of value, a hedge against fiat debasement — then this correlation data is your thesis walking to the gallows. A 95% correlation to equities means Bitcoin is not a hedge. It's a leveraged tech stock with worse liquidity and no earnings.
The institutional adoption that Bitcoin's bulls celebrated — ETFs, corporate treasuries, sovereign wealth fund allocations — has achieved exactly what it always would have: it made Bitcoin a correlated risk asset because it's now held by the same people who hold the S&P 500, rebalanced by the same portfolio models, and sold in the same liquidation cascades.
This doesn't mean Bitcoin goes to zero. It means the intellectual framework that justified its price appreciation — scarcity, uncorrelatedness, monetary revolution — has been empirically falsified. What's left is pure momentum trading and network effects. That's a legitimate game. But it's not the game the believers signed up for.
The Agentic AI Delusion
Singapore's hottest job skill is now "agentic AI workflows." Job postings are rewriting themselves. Recruiters are scrambling. The narrative is that AI agents will handle complex multi-step tasks autonomously, freeing humans for higher-value work.
Here's what the practitioners are actually saying: "I still work 40 hours a week. AI turns it into 120." And another: "AI won't tidy your business. It just moves the mess faster."
The gap between the agentic AI narrative and the agentic AI reality is the same gap that existed between "the cloud will simplify IT" and the reality of cloud cost management nightmares. Agents don't reduce work. They multiply the number of things you can attempt simultaneously, which is fundamentally different from reducing effort.
Visa's AI agent payment pilot with Bill and Billtrust is telling: the problem they're solving isn't "AI can't pay invoices." It's "shared login credentials are a security nightmare." The AI framing is marketing. The actual innovation is credential management.
The Bottom Line
The dominant narrative of October 2026 is that AI is the new industrial revolution, and the numbers are getting bigger every day — $750B in compute, $5.6B acquisitions, $325M rounds for aviation AI. But underneath the capital flows, three structural problems are becoming impossible to ignore: the return-on-investment question for AI infrastructure spending, the failure of Southeast Asia to convert digital adoption into AI value capture, and the quiet erosion of the geopolitical assumptions that made US tech dominance seem permanent.
The market is pricing in a future where AI spending compounds returns the way mobile internet did from 2010-2020. But mobile internet had a 10-year runway of cheap capital and genuine consumer surplus before anyone demanded profitability. AI is demanding ROI now, while simultaneously requiring capital expenditure at a scale that makes the mobile era look like a lemonade stand.
The $750B isn't a bet on technology. It's a bet on time. And time, unlike compute, doesn't scale.