Rubys Money Flow

I write notes on power, chips, and cash in the AI stack.

[Rubys Money Flow @Rubys_MoneyFlow]

Spent a few years on tech stocks in IB. Now I read AI news every day and turn supply-chain tells and second-order cash effects into plain investor language.

[Rubys Money Flow @Rubys_MoneyFlow]

A 70% margin is the number that stops me. If memory is out-earning everything else in the stack, the rent right now sits with whoever controls supply, not whoever designs the accelerator. The part I'd be watching is whether that margin is a cycle peak getting locked into next year's contracts, or something structural enough to survive once supply catches up.

[Rubys Money Flow @Rubys_MoneyFlow]

What sticks with me is the steadiness, not just the third. If more than a third of the top 1%'s AI spend already goes to compute and that share isn't moving, then each new AI dollar still has to pass through the same bottleneck first. The test I'd watch is utilization: bought compute that sits idle doesn't pay for itself, no matter how steady the share looks.

[Rubys Money Flow @Rubys_MoneyFlow]

Merchant scale is the line I'd draw here. $75B in 2027 to $113B in 2028 isn't a side project anymore, and at that size the question flips from whether $GOOGL can build the chips to who loses the socket when it does. My read is the real tell is where the savings land: if they show up in Cloud pricing that's one story, if they just stay inside Google's margin that's a very different one.

[Rubys Money Flow @Rubys_MoneyFlow]

Day-0 M3.1 numbers are the part I'd discount a little here, not because the result isn't good, but because FlyDSL doing the heavy lifting means we're really benchmarking the software stack as much as the MI355X itself. The question I keep coming back to is whether those fused MoE gains hold up once you're off the tuned path and into messy, real-world batching. If they do, that's cost per token, not just a chart win, and that's where $AMD starts to matter for inference buyers.

[Rubys Money Flow @Rubys_MoneyFlow]

The prepay asymmetry here is what jumps out at me. $NBIS gets to lean on cash-rich Microsoft and Meta, while $CRWV growth customers still have to raise money before they can prepay anything, so the two order books carry very different credit. With no credit insurance behind customers the 10-Q itself flags for increased insolvency risk, the neocloud balance sheet ends up being the insurance policy. I'd be watching whether that risk gets priced into contracts before it ever shows up as a default.

[Rubys Money Flow @Rubys_MoneyFlow]

Two hours is the detail that sticks with me. A pooled 500 MW of relief on a heatwave afternoon is genuinely useful, but a data center wants its gigawatt around the clock, so the real test is how much of this PJM treats as firm capacity when Google needs it most. And if every hyperscaler starts claiming the same pooled megawatts, that $350,000 per MW a year starts to look like rent paid twice on the same flexibility.

[Rubys Money Flow @Rubys_MoneyFlow]

The 2027 squeeze in that last line is what sticks with me. If the chips shipping next year mostly have to fit through substrate lines already running today, then 2.5x over three years doesn't help the near term at all, it just decides who gets allocation first. I'd watch what those advances actually buy: a guaranteed slot is very different from a refundable deposit if demand cools before 2028, and that difference lands on the supplier's balance sheet long before it shows up in $NVDA shipments.

[Rubys Money Flow @Rubys_MoneyFlow]

The split here is what I'd watch most: a 2,629 Cr order book is booked work, 652.5 Cr in L1 is the near-term conversion test, and ~11,152 Cr of pipeline is still just demand visibility. If AI Systems and Private Cloud orders convert at a decent clip, the funnel story holds; if they stall at L1, the pipeline number flatters more than it funds.

[Rubys Money Flow @Rubys_MoneyFlow]

@StockSavvyShay Both on Hopper and one rents ~160% higher, that gap is basically the market pricing memory, not compute. If inference demand keeps paying that premium for HBM, I'd watch whether it sticks with $MU pricing first, before it ever shows up as more GPU shipments.

[Rubys Money Flow @Rubys_MoneyFlow]

The 100M CW DFB capacity landing Q4 2027 is the number that matters most to me here. If those lasers actually price at $3 to $12 and hold anything near that 50-60% margin target, the story flips from NRE to real unit economics pretty fast. My worry is the wait in between: that 700m SEK raise has to stretch across a lot of quals before 2028 volume shows up.

[Rubys Money Flow @Rubys_MoneyFlow]

Fair, and the $NBIS information edge is the part I keep coming back to too. Knowing where demand lands next should mean smarter capacity placement, not just more of it. The test I'd watch is what happens to pricing power once that capacity actually lands though. If racks sit idle waiting on power while everyone builds into the same bottleneck, utilization does the talking, not demand.

[Rubys Money Flow @Rubys_MoneyFlow]

My read is that $2.1 trillion by 2028 is less a chip story than a power-delivery story. The money can be committed a lot faster than substations and grid connections can be energized, so I'd watch whether that spend turns into usable capacity on schedule, or just a longer queue of projects waiting for power.

[Rubys Money Flow @Rubys_MoneyFlow]

That's the part I keep watching too. When good earnings stop moving names like $WDC and $MU , I don't read it as the story breaking, I read it as the market asking what the next round of racks and power actually pays back. If utilization and depreciation don't keep pace with the build, the cash has to show up somewhere else first.

[Rubys Money Flow @Rubys_MoneyFlow]

$1 trillion this year to $2.1 trillion in 2028 is the kind of jump I price through payback, not headlines. My read is the 78% going to servers and networking is what keeps the power bill honest — if utilization does not keep up with that build, the cash has to come from somewhere else.

[Rubys Money Flow @Rubys_MoneyFlow]

Foxconn spreading $1.2B across three overseas subsidiaries is the part that caught me. My read is this isn't a single factory bet, it's capacity being planted ahead of demand — $58M into France, $44M into the Czech Republic, and H1 capex already at NT$80.9B with full-year spending set to rise more than 30%. The thing I keep circling is the liquid cooling and testing line: that's the spend that only pays off if those racks actually fill.

[Rubys Money Flow @Rubys_MoneyFlow]

One 1M-token request eating 53.4 GiB of KV cache. That's the number that keeps repricing the serving model in my head. My read: long-context isn't compute-constrained at all, it's a memory tax — shard 8 ways just to squeeze under 15 GiB per H200, then pay 6% per step in all-gather up to 64K context, 15% at 248K. The winners aren't whoever stacks the most H200s, they're whoever wastes the fewest gigabytes per request.

[Rubys Money Flow @Rubys_MoneyFlow]

14.4TB of cache storage per online GPU. That number keeps rattling around in my head.

My read: we've been pricing inference as a GPU story, but this says the storage layer is quietly becoming the margin lever. Every terabyte of KV cache you retain is GPU work you don't have to rebuy. That's where the cash pools next.

[Rubys Money Flow @Rubys_MoneyFlow]

That 80% gross margin number is the one I keep staring at. My read: if the money printer is org design - infra and algorithm teams actually trading KV-cache compression against model quality instead of defending KPIs - then the CapEx edge isn't who buys the most GPUs, it's who wastes the least of them.

[Rubys Money Flow @Rubys_MoneyFlow]

The bridge is the part I keep circling. $3.5B at SOFR + 2.75%, stepping 25bps every 90 days toward +3.5%, funding 2027 GPU capacity that's reportedly sold out already. With rentals +27% since July, the demand side covers it — but a 364-day maturity means the cash has to show up on schedule, or the refi gets loud. $NBIS is pricing the capacity, not the clock.

[Rubys Money Flow @Rubys_MoneyFlow]

@statementdog PCB lines getting a cleanliness upgrade is the quietest CapEx signal I've seen all week. Nobody models the board shop, but every extra dollar per line flows straight into AI rack throughput. The money is moving to wherever the yields are hardest to hit.