Over the past week, the AI hardware news I’ve been tracking adds up to more than $610 billion in capital deployed globally — in just seven days.
Not valuations. Not market cap. Actual capital expenditure commitments. Korea $550B, Japan $6B, Qualcomm’s new accelerator, Kawasaki Heavy Industries’ $1B AI infrastructure bond — this round of moves has already surpassed the wildest half-year of the 2000 dot-com bubble in scale. But this time the money isn’t flowing into web pages. It’s flowing into chips, memory, and power. Watching all of this over the past few days, I’ve been thinking: for investors and for builders like us making products on top of AI, what does this gamble actually mean?
The Real Story Behind AI Training Bottlenecks: From GPU Scarcity → Memory Scarcity → Power Scarcity
Honestly, everyone watches AI through the lens of models, but the real bottleneck was never the models — it’s been the hardware.
From 2023 to 2025, the bottleneck shifted from GPU scarcity to memory scarcity, and is now pushing toward power scarcity. When GPUs were tight, everyone scrambled for H100s and NVIDIA raked it in — but the part that actually throttled the H100 wasn’t the GPU core, it was the HBM high-bandwidth memory. On the B200, the HBM3E stacked on top has its capacity locked up entirely by NVIDIA at SK Hynix, while Samsung is chasing hard but its yields can’t keep up.
That’s why South Korea just committed $518B to build 4 memory fabs plus $52B for the central regions, totaling $550B (TechCrunch). This isn’t just about filling upstream capacity — the key is that Samsung + SK Hynix are trying to flip themselves from being NVIDIA’s downstream suppliers into becoming the dominant players in AI hardware.
Why did downstream hardware investment kick off so late? Because for the past two years people were still watching and waiting to see if “this AI hype cycle would cool down again.” By 2026, GPT-6, Claude 4, and Gemini 3 are all live, inference costs have come down, user numbers are real — only then did people dare to bet on 10+ year capacity. Japan is also putting in $6B to back SoftBank-led AI model development (Nikkei), and Kawasaki Heavy Industries issued a $1B bond to dive into AI infrastructure. East Asia is pushing “AI sovereignty” as a national-level agenda.
Qualcomm’s Challenge: What Bypassing HBM Actually Means
Then there’s Qualcomm — and this move is interesting.
Qualcomm’s new AI accelerator series announced in early July leads with “HBM-independent” (Qualcomm), going with an LPDDR + on-chip SRAM combo. Sounds like a compromise, but it’s actually a completely different path.
If this path works out, the implications for the ecosystem run three layers deep.
Layer one: part of NVIDIA’s tech moat gets bypassed. The H100/B200 doesn’t just rely on CUDA — it’s the entire locked-in combo of CUDA + HBM + NVLink. With HBM bypassed, the logic of “you must buy NVIDIA for AI inference” loses a corner.
Layer two: more subtle implications for TSMC. NVIDIA’s B200 is on TSMC 4nm, Qualcomm’s new accelerator is also on TSMC, and Samsung/SK Hynix’s HBM packaging has to go through TSMC for CoWoS. No matter who wins, TSMC takes orders this round — and that’s why I think it’s a better medium-to-long-term bet than NVIDIA.
Layer three: the most tangible impact is for people building AI applications. If Qualcomm can push inference costs down to a third of NVIDIA’s at the same tier, there’s a good chance the OpenAI/Anthropic API costs we’re paying now drop by half within six months — and every product idea shelved because “the API is too expensive” will get pulled back out and rebuilt.
4 Practical Judgments for Investors and Builders
So I’ve boiled it down to four judgments.
One: don’t chase NVIDIA’s short-term highs. The lesson from last round’s $600B+ market cap evaporation in a single day hasn’t been fully digested yet. The valuation has already priced in three years of flawless execution — miss any milestone, and the pullback won’t be small.
Two: TSMC benefits on both sides. Building HBM packaging for Korea, B200 for NVIDIA, the new accelerator for Qualcomm — whoever fights whom, TSMC’s still collecting the orders.
Three: AI application developers don’t need to wait for the hardware reshuffle. The product you’re building right now on top of OpenAI/Anthropic APIs — whoever swaps in underneath is transparent to you. Migration costs are near zero. The people who wait are the ones who miss the window.
Four: over the next six months, watch two signals — power (US state-level data center power permits, Japan/Korea nuclear restart progress) and the HBM4 production timeline (2027 Q2 is the key milestone).
Closing: Capital Flows Out of Bubbles, Infrastructure Stays
This is the largest single-sector global capital deployment since the dot-com bust. There’s some bubble in it, but the underlying demand is real. Money will flow out of the bubble, but when it does, the infrastructure will genuinely remain — unlike back in 2000 when all that was left was dark fiber no one used. What this round leaves behind is the foundation AI will run on for the next decade. Every product I’m building is a bet on that assumption.