Vaporwave AI capex hero
HERMES DISPATCH · №10 · 30 JUL 2026

The Layoffs
ARE The Capex Plan.

▼ SWIPE UP
THESIS01 / 09
The confession beneath the headcount

Big Tech isn't paying for GPUs with revenue anymore.
They're paying with payroll.

Two headlines this month, one story:

Oracle cut 21,000 people (13% of the company), booked $1.84B in severance, then re-directed the cash toward its $300B compute contract with OpenAI.

Amazon is now ~30,000 corporate roles lighter since late 2025 while committing ~$200B in 2026 capex, mostly AWS/AI silicon.

Read the mechanism, not the memo.

THE LEDGER02 / 09
Read the basement, not the blog post

The 2026 capex bill is $600–725B. Someone has to pay.

21,000 Oracle staff cut in FY26 · $1.84B severance · re-routed toward the $300B OpenAI compute deal (Yahoo Fin, 25 Jul)
~30,000 Amazon corporate roles removed since late 2025 · ~$200B 2026 capex, most of it AWS + AI silicon (Barchart / IBTimes, 22 Jul)
$725B MAG-7 hyperscaler capex projected for 2026 · Reuters (17 Jul): investors already positioning for slower spend growth in 2027

The math: revenue can't fund the buildout inside a quarterly cycle, so the payroll line pays for the compute line. Layoffs are the financing.

WHAT THE DATA HIDES03 / 09
Tomato-soup test · the average buries the story

"AI is causing layoffs" is a lie of averages.

The clean stat is TechJack's tracker: AI-cited layoffs jumped from 0.6% (2024) → 4.5% (2025) → 25% in July 2026. What's under it:

AProductivity cuts — the CEO story: "AI made these roles redundant." Real, but a minority.
BCapex-funding cuts — Oracle, Amazon. The role wasn't automated. It was traded for GPU allocation.
CBusiness-cycle cuts wearing AI clothes — TechCrunch (25 Jul) lists 20+ names citing "AI" for what looks a lot like normal restructuring in a slowing economy.

If you're an operator, only bucket A is a signal about YOUR business. B and C are signals about theirs.

THE GRAVEYARD04 / 09
A real AI faceplant · transferable lesson

The moment capex met the grid.

HEADSTONE · 22 JUL 2026 · WISCONSIN

Oracle vs. We Energies · $7B collateral standoff.

Oracle went to court to fight a Wisconsin PSC ruling that would require it to post $7 BILLION in collateral before the utility builds power for its ~1 GW Port Washington data centre.

The utility isn't gouging. It's pricing the risk that a hyperscaler walks away mid-buildout and leaves ratepayers holding a stranded power plant.

Lesson for us: the compute race is now a power race, and power lives inside 4-year utility-approval cycles. Any AI-native ambition — Alaya's, Hyro's — that assumes compute is elastic is pricing off the model card, not the utility bill. The bottleneck is a megawatt, not a GPU.
WHAT AI'S GOOD AT · AND NOT05 / 09
Honest capability map · Q3 2026

Kill hype in both directions.

✓ Actually compressing
Structured back-office work with lots of forms. Broker file assembly, wholesale quoting, first-draft memos, code refactors, meeting → task extraction. Real productivity, real payback.
✗ The hype is still bluffing
Autonomous strategic judgement. VentureBeat's own Jul-2026 wave: 50% of enterprises shipped an agent that passed eval, failed a real customer. Klarna is rehiring 700 humans. The pattern is agents that test well and trade poorly.
→ Real edge in H2 26
Proprietary workflow context, not more compute. With tokens crashing (see Muse Spark, Dispatch #08), the moat isn't "we have GPT-5.6." It's we have the FMCG-ops or mortgage-broker workflow data nobody else has. That's the last defensible mile.
THE OPERATOR PLAY06 / 09
Three second-order moves this quarter

Play the aftermath, not the party.

1
Buy the discount, don't finance the buildout. Model prices are falling 40–75% YoY. Lock 12-month usage on the cheapest capable model tier NOW. You're the buyer of a fire sale; act like one.
2
Vendor concentration risk = walkaway risk. If one hyperscaler is >40% of your critical stack, you're now correlated with Wisconsin power hearings. Dual-provider your top three AI-dependent workflows this quarter.
3
Build the workflow moat before it commoditises. Every clean piece of proprietary data (broker file, wholesale order history, packaging spec) is worth more each week tokens get cheaper. Data-cleaning is the highest-ROI internal AI project in the shop right now.
THE INVESTOR STORY07 / 09
For the US raise · investor room framing

Be the org that gets more valuable when tokens hit zero.

The bear case in the room right now: "MAG-7 is spending $725B and Cahn says they need $3T in revenue to justify it. Where's the return?"
The Hyro answer isn't "we use AI." It's this: "We're Customer Zero for the AI-native FMCG operating system. Our proprietary FMCG-ops data appreciates as model costs fall. We're the compounding asset on the demand side of the fire sale."
Same for Alaya: every broker workflow captured is octopus-limb IP — cloud-native, guarded, humans-in-loop. The tenants of the compute economy get squeezed. The owners of workflow context get repriced upward.

Otto's frame from Dispatch #06 still holds: "the AI of running FMCG could be worth more than Hyro itself." The layoff cycle is proving it — labour is the variable that gets swapped for compute, but curated workflow data is what makes the compute worth anything.

SIGN-OFF08 / 09
"Read the basement, not the blog post."

Reply-window question for you, Otto: Alaya's payroll is a smaller line than Oracle's, but the same principle applies — which broker workflows do you keep cash-funding with headcount because the AI still fails the customer, and which are you already ready to trade payroll for context? Curious where your bucket A / B / C split lands.

— Steve

Dispatch №10 · 30 JUL 2026 · Built by Hermes
Sources this issue: Yahoo Finance, Reuters, TechCrunch, JPost, Barchart, IBTimes, Motley Fool, Capacity Global, TechJack, Basic-Tutorials — all posts published this month.