TMT Brief · Wednesday, September 23, 2026

Koko: AWS just committed to buying more than three million additional Nvidia GPUs before twenty twenty nine. Amazon has one point five eight million employees. Per TechRadar, Amazon will soon have more GPUs than staff.

Max: [excited] That's not a metaphor, that's a literal ratio now. Chips outnumbering people on the payroll.

Koko: And it's stacked on top of the H100s and H200s they already run. This is the first time hyperscaler capex has had a number this concrete attached to it.

Max: Fast around the rest. Nvidia put one point five billion into SoftBank's SB Energy ahead of its IPO, on top of one hundred five billion in residual-value guarantees tied to four point two five gigawatts in Ohio.

Koko: IREN's latest call disclosed four billion in contracted ARR against nineteen billion in funding, and sixteen billion of that is debt-like. Qualcomm shipped an AI Android chip right as a memory shortage squeezes smartphone costs.

Max: And Alibaba's new AI chip and data center plans moved its own share price. Infrastructure disclosure is now a market event, not a footnote.

Koko: Here's why this week matters for the board deck: the financing stack behind AI, vendor guarantees, prepayments, GPU-backed debt, is now doing real balance-sheet work. Enterprises are still overspending on tools that haven't proven ROI.

Max: The infrastructure side is getting more sophisticated by the week. The usage side is still guessing. That gap is the whole show.

Koko: Let's sit with the AWS number, because three million incremental GPUs before twenty twenty nine is a commitment, not a wish list. That's years of capacity locked in.

Max: Here's my read: this is discipline dressed up as aggression. Amazon's locking in supply and pricing now, before the next wave of demand makes it more expensive.

Koko: I'd push back on that. Locking in three million chips before you know the shape of enterprise demand isn't discipline, that's a bet the whole market keeps growing at this rate.

Max: Fair, but look at what Nvidia's doing alongside it. One point five billion into SB Energy, one hundred five billion in residual-value guarantees in Ohio. Nvidia is financing its own demand curve.

Koko: Which should make any CFO nervous. When your biggest supplier is also underwriting the power plants and the customers buying its chips, the whole chain is circular.

Max: [chuckles] Circular, but not necessarily fake. IREN's numbers give you the other side of that coin. Four billion in contracted ARR against nineteen billion raised, sixteen billion of it debt-like.

Koko: That ratio is the one to benchmark. If your AI infrastructure play looks anything like IREN's, your CFO needs to know how much of that capital structure is prepayments and converts versus actual equity cushion.

Max: The action item isn't 'is this bubble or not.' It's: go find your own version of that sixteen-to-three ratio and know it cold before the next board meeting.

Koko: Agreed on that much. Where we disagree is whether AWS ordering three million GPUs is prudent supply-chain management or the industry's biggest single bet yet.

Max: Both things can be true. That's what makes it uncomfortable.

Koko: Qualcomm shipped a new Android chip built for AI this week, per CNBC, right as a memory shortage is squeezing smartphone component costs.

Max: Timing that nobody planned for. You launch your flagship AI silicon into the one input market that's suddenly tight.

Koko: Memory has quietly become a live margin line, not a rounding error in the bill of materials. Every device maker running AI features on-chip needs more of it.

Max: This is the input inflation story CFOs haven't fully priced. Everyone's been watching GPU cost and power cost. Memory just jumped the queue.

Koko: And it hits differently depending on where you sit. Component buyers eat it in cost of goods. Device makers either pass it through in price or absorb it in margin.

Max: Given the smartphone market's pricing sensitivity, I'd bet on margin absorption first, price increases second, and only after competitors move.

Koko: That's the safer bet, but it's also a slower bleed. Quarter over quarter, that's real basis points off hardware margin while everyone waits to see who blinks.

Max: Worth flagging to the CTO and supply chain lead now, not at the next earnings call. If memory stays tight into next year, component sourcing strategy needs to move before the shortage becomes the excuse.

Koko: McKinsey data cited by CIO Dive says most enterprises are overrunning their AI budgets specifically on coding agents, despite weak measured ROI.

Max: We've flagged this gap before, but the fact that it's concentrated in coding agents specifically is the new detail. That's the one category everyone assumed had obvious payback.

Koko: That's exactly why it's uncomfortable. If the easiest use case, developer productivity, is overrunning budget without proof, the harder use cases have no excuse.

Max: Devil's advocate: maybe the ROI measurement is the problem, not the spend. Developer productivity is notoriously hard to quantify in a clean dollar figure.

Koko: Sure, but that's not a reason to keep the budget open-ended. That's a reason to build the measurement before you approve the next tranche of spend.

Max: This is squarely a CFO and CIO conversation now. Chargeback, usage tracking, some way to tie coding-agent licenses to actual velocity, not vibes.

Koko: And it connects straight back to the pricing story. Frontier model costs have been falling all year, so if your coding-agent bill is still climbing, that's a usage problem, not an input-cost problem.

Max: [chuckles] Which is a much less comfortable conversation to have with the engineering org than 'prices went up.'

Koko: Zoom out and three things are moving together. Infrastructure financing is getting more creative, and more circular, by the week.

Max: Vendor guarantees, GPU-backed debt, hyperscaler-to-supplier equity stakes. The stack is doing more work than a simple capex line can show.

Koko: Input costs are showing up in places nobody was watching. Memory, not just power or chips.

Max: And enterprise AI spend keeps outrunning the proof, even in the category everyone thought was safest.

Koko: None of these are new stories individually. Together they're one story: capital is getting smarter, usage discipline isn't keeping pace.

Koko: So what's a board-deck item, and what's a call to make this month. Start with the quarter-out planning: benchmark your own capital structure against IREN's ratio, debt-like financing versus equity, before your next infrastructure commitment.

Max: That's a next-quarter conversation with the board, because it means actually mapping every prepayment and GPU-financing instrument you're carrying.

Koko: The move-this-month item is coding agent spend. Get the CIO in a room, put chargeback and usage tracking on every AI tool license before the next renewal, not after.

Max: And watch memory costs specifically. If you make or buy AI-enabled hardware, get ahead of pricing decisions before the shortage becomes the story you're reacting to instead of managing.

Koko: What would change our answer? If enterprise ROI data on coding agents flips positive next quarter, the spend conversation changes completely.

Max: Until then, treat every AI tool line item as unproven until it's measured, not because AI doesn't work, but because right now, nobody's shown their math.

Koko: Full write-up's at kokoknows dot A I.

Max: Here's the open question for the coming weeks. Does memory pricing become the next capex headline, the way power did this year?