Tech moves fast, but you're still playing catch-up?

That's exactly why 200K+ engineers working at Google, Meta, and Apple read The Code twice a week.

Here's what you get:

All delivered twice a week in just 2 short emails.

Join 200K+ engineers

Most Meta writeups split two ways. Bulls say "AI is monetizing the ad stack, buy it." Bears say "capex is out of control, sell it." Both are vibes until you put them next to a cash flow statement.

I'm a controller. With NVDA the question was whether the customer's wallet stays open. With META, the question is whether META's own wallet can absorb what management just guided. Different question. Same discipline. And the print on April 29 will resolve a chunk of it in 72 hours.

Here's where we are.

Q4 2025 (December quarter):

Full year FY25: revenue $200.97B, +22%; net income $60.5B; free cash flow $43.6B; capex $72.2B; share repurchases $26.3B; dividends $5.3B. Cash and marketable securities $81.6B.

The Family of Apps business is a machine. ROIC 27.8%. ROE 30%. They printed $43.6B of free cash and returned $31.6B to shareholders. Nothing in this row is the problem.

The problem is the next row.

CFO Susan Li guided FY26 capex to $115–135B — up from $72B in FY25. Midpoint $125B is +73% YoY. Capital intensity moves from 36% of revenue to roughly 50%, which is a number you don't see outside utilities, telecoms, and oil majors mid-build-out.

Where it goes:

That last bullet is the tell. A company building gigawatt data centers is also writing $50B+ in third-party cloud checks. Either internal build slipped or demand outran the construction schedule. Probably both. Near-term cash goes out the door faster than the data centers come online.

Total expense guide for 2026: $162–169B, vs $117.7B in FY25. That's +40% expense growth against revenue guided to grow ~25%. The math doesn't have to be told to you.

Before the cash flow math, ask the harder question: how much of the $125B is required spend to keep the ad business from degrading, versus new bet spend on AI surfaces that don't exist yet?

My read: roughly 60% defensive, 40% offensive — and trending defensive.

Defensive capex isn't optional. It's table stakes. The GEM ranking model that lifted Facebook CTR ~3.5% is a defensive win — without it, the same number degrades against whatever Google and TikTok are spending. Llama training is defensive: not having a frontier model means renting reasoning from OpenAI on someone else's roadmap. Post-ATT first-party AI inference for ad targeting across 3.58B daily users is not a luxury — it's the cost of staying in the auction. Most of the $35B CoreWeave bill is Llama training. Most of Hyperion is general inference, the lion's share serving existing ad and feed surfaces. MTIA silicon is primarily inference — defensive against the NVDA tax.

Offensive is the smaller slice. Advantage+ as a standalone revenue product. Click-to-WhatsApp business AI. Meta AI monetization (not yet). Glasses as a hardware platform.

The trap most takes fall into is treating the full $125B as "the AI bet." It isn't. Maybe $50B is the bet. The other $75B is what you have to spend to not lose. That changes the IRR math entirely. If defensive capex earns roughly the cost of capital — call it 9% — the offensive 40% has to clear 15–20% to make the package compound at all. High bar. But that's exactly what infrastructure investment is supposed to do.

The signal that matters isn't capex level. It's whether ad revenue accelerates alongside the spend. Revenue stuck at +22–25% while capex doubles means defensive is failing — paying more to stay in place. Revenue moving to +28%+ means offensive is starting to earn. Q1 was guided to +31% on FX tailwinds. The cleaner read comes in Q2 when FX rolls off.

This is the part most "META is cheap on forward earnings" pieces skip.

Take consensus 2026 revenue at $250B, expense guide of $165B → operating income near $85B. Tax it at ~18%, net income lands near $70B. EPS around $30. Forward P/E ~22x. Looks fine.

Now run cash flow. OCF scales with net income plus depreciation — call it $135–140B. Subtract capex of $125B. Free cash flow lands between $10–20B depending on working capital. That's a 50–80% drop from FY25's $43.6B.

Then look at capital return. $26B of buybacks and $5B of dividends in FY25. If management holds buybacks flat in 2026, they're funding shareholder returns out of the balance sheet, not free cash. They have $81.6B cash; they can do that for a while. But this is the year capital allocation logic flips. If the AI investment cycle pays back, FCF re-accelerates in 2027–2028 and the buyback math is fine. If it doesn't, buyback gets cut or debt expands to fund it.

That's the trade. Earnings look fine. Cash looks bad on purpose. Management is telling you the IRR on $125B of AI infra beats the IRR on buying back stock at $675. Maybe. The data to grade that thesis isn't here yet.

The 8,000 layoffs starting May 20 — 10% of the workforce — are management telling you they see the same expense problem. Cost cutting at a 24%-growth company is a tell.

Cash flow compression is one risk. The other is that the build itself doesn't deliver.

Custom silicon is the right bet, not a guaranteed one. Second-sourcing against NVIDIA is necessary — Meta would be irresponsible not to try. But CUDA is the real moat in accelerated compute, not the chips. Switching costs across a multi-year ML training stack are real and ugly. Performance per watt still favors NVIDIA at the frontier. Meta is betting it can vertically integrate fast enough to matter before its 2027 capex plan goes stale. That's not proven. And milestone slips don't show up as a clean kill signal — they show up as another quarter of $35B+ rented compute and another year of NVDA tax.

Power is the gating constraint, not steel. The Hyperion campus alone required Meta to underwrite seven new natural gas plants and ~240 miles of 500-kV transmission with Entergy Louisiana. Multiply that across the broader build. Grid interconnect timelines are extending; build costs are rising non-linearly. Every quarter the internal build slips, the external compute bill grows. The CoreWeave commitment ballooning from $14.2B to $35.2B in six months is what that loop looks like in real time.

This is a timing problem, not a demand problem. Demand for AI compute isn't the question — the question is whether Meta's capacity comes online during the highest-IRR window or after it. If the hyperscaler cohort overbuilds, pricing power shifts to the AI labs and consumers, and the math on $125B/year of capex looks very different. That's how infrastructure cycles end — not with dramatic demand collapse but with supply catching up six months too soon. None of that is visible in 2026 numbers. It's visible in 2028 depreciation against 2028 revenue, when there's nothing left to do about it.

Q4 2025: $6.0B operating loss on $955M revenue. Full year: $19.2B loss on $2.2B revenue. Cumulative since 2020: roughly $80B in losses.

The pivot is real but slow. Roughly 70% of RL spend now goes to AI glasses and wearables, 30% to VR/Quest. Ray-Ban Meta sales tripled in 2025. The $799 Ray-Ban Display launched in September and sold out of initial inventory. EssilorLuxottica capacity ramps to 10M units/year by end of 2026. Quest revenue declined 12% in Q4 because they didn't ship a new headset in 2025.

Zuckerberg called 2026 the peak year for RL losses. I don't trust that forecast — every prior peak guide on RL has been blown through — but the pivot from VR-first to wearables-first is the right call. Smart glasses are a real product with real demand. Quest never crossed that threshold.

Subtract the $19B RL drag and the rest of Meta runs at closer to a 50% operating margin. The ad business is even more profitable than the headline suggests — RL is just a $19B annual donation to a long-dated bet.

For 2026, assume RL stays flat to slightly worse. Don't model improvement; let management surprise you.

The reason capex is even thinkable is that the ad business is gushing.

Per Emarketer, 2026 is the first year Meta's net ad revenue exceeds Google's: $243.5B vs $239.5B. Meta growing 24.1%, Google 11.9%. Five years ago iOS privacy changes were supposed to cripple Meta's targeting and TikTok was going to gut its engagement. Both threats arrived. Both got absorbed. The ad machine kept compounding.

Competitive picture in 2026:

Meta's edge in 2026: first-party data is back in style. With cookie deprecation and ATT in the rearview, advertisers want platforms with logged-in users at scale. Meta has 3.58B daily users. That's the moat.

Sell-side average target ~$839 maps to my base case. The setup pays you +28% in the base, hurts −11% in the bear. Asymmetric upside, but not the screaming +46% asymmetry NVDA had at $208. META at $675 is closer to fair than dislocated.

Two of these flipping is enough to cut size:

None of those are narrative calls. They're prints on a calendar. Three of them grade in 72 hours.

Long, sized lighter than NVDA. The thesis is durable; the next four quarters are noisy.

Methodology check:

I'd rather own the next 12 months at $675 than the past 12 months at $796. The valuation reset since September did some of the work. The remaining work is execution.

Three ways to play it into the print:

What I won't do: size this like a high-conviction revision-momentum trade. Revision breadth is positive but mid-pack. Capex absorption is the swing factor and the data to grade it lands quarterly.

Two of three break, position size comes down before consensus moves.

What ends this trade isn't competition. It isn't regulation. It's a CFO sitting on the call Wednesday saying "we now expect 2026 capex in the range of $130–145B." That sentence prices in 30 minutes, not 30 days.

Until then: starter or zero. Watchlist active. Eight numbers checked Wednesday after the close.