MSFT and META reported within an hour of each other Tuesday night, both defending the same thing: an AI infrastructure bill so large it is now bending their cash flows out of shape. Both quarters, once you get underneath the headline, were fine. And by the next day’s close the market had sent Microsoft up about 14% and Meta down about 9%, a 23-point gap between two companies telling substantially the same story. The tidy version making the rounds is that Microsoft beat and Meta missed, so the market bought one and sold the other. The prints don’t support that, because both headlines were misleading and the people who move these stocks knew it within minutes. What actually split them was the forward trajectory the quarter revealed: whose core metric is accelerating, whose spending funds itself, and who would show enough to prove the buildout is working. On all three the two companies diverged hard. Read the footnotes and a more uncomfortable point shows up underneath: both are quietly telling you less than a year ago, at exactly the moment the spending makes disclosure matter most. Microsoft just had more credibility to spend, so it got away with it.Both headlines lied, in opposite directionsStart with the numbers that ran the coverage, because both were misleading. Meta missed. GAAP EPS of $6.18 came in 14% below the $7.19 consensus, and the screens traded it as a miss within seconds. But $2.4B of legal charges and $1.18B of severance, both disclosed, both one-time, account for almost the entire gap. Strip them at the quarter’s own tax rate and core EPS was about $7.35, a beat, and management said as much in plain language: excluding those items, operating income grew 9% year over year. Revenue beat and grew 28%. The business the headline called a miss was, on the numbers management handed the Street, a modest beat.Microsoft crushed it. GAAP EPS grew 32%. But a large chunk of that came from a line that has nothing to do with selling software: its stakes in OpenAI and Anthropic. The 10-K discloses a $6.5B net gain in other income from those holdings this year against a $4.8B net loss last year, an $11.3B swing that the press release quietly folded into a $3.2B Anthropic gain discrete-item mention. Strip the equity-method noise and the cleaner, ex-OpenAI EPS growth was about 22%, not 32%. Still very good, just not the number on the wire.Here is the tell that the headline wasn’t the explanation: it was misleading in opposite directions, and the analysts who cover these two names strip out one-time charges and equity-method marks in their sleep. If the market had simply traded the print, Meta would have bounced once the ex-items math circulated and Microsoft would have faded once the OpenAI marks came out. Instead both moves held and widened into the next session. The market wasn’t reacting to the quarter that just happened. It was repricing the quarter that hasn’t happened yet, and on that score the two companies genuinely diverged.What the market actually pricedStrip the headlines and three forward-looking things moved, all in Microsoft’s favor and all against Meta. The first is the second derivative on the metric each stock lives by. Azure grew 43% in constant currency and management guided the next quarter to about 45%, acceleration on top of acceleration, while commercial bookings excluding OpenAI more than doubled their growth rate, 7% to 18%. Meta’s core metric went the other way: ad impressions grew 14%, down from 19% a quarter earlier, with price per ad flat at 12%, so the growth is coming more from charging more than from showing more, a shift third-party panels confirmed was sector-wide. One engine is speeding up. The other is quietly changing its mix in a way that is harder to sustain.The second, and this is what best explains the size of Microsoft’s move, is that the print killed the specific bear thesis that had been holding the stock down. Microsoft went into the quarter off 8% over three months against a rising market, on one fear: that AI capex was crushing cloud margins. Then Microsoft Cloud gross margin printed 65%, above the roughly 64% the company itself had guided in April, while Azure accelerated. The exact thing the market had been marking the stock down for did not happen. A 14% move isn’t enthusiasm about an earnings beat, it’s the market retiring a thesis it had been charging the stock for since spring. Meta offered no such relief, because its margin problem isn’t an accounting artifact, it’s that operating expenses are growing at more than twice the rate of revenue: R&D rose 67% year over year, near 58% excluding severance, against 28% revenue growth, with no sign on the call that the pace is about to break. That is forward margin compression, and pricing it now rather than later is the correct call, not a dumb one.The third is what each was willing to promise about the spending. Microsoft paired its guide-above-consensus with a $678B contracted backlog. Meta, asked point-blank for even directional 2027 capex, declined, while its free cash flow was collapsing and it had suspended the buyback to conserve cash. One company bounded its commitment and showed the demand underwriting it. The other left the biggest number in its model open-ended while the cash to fund it visibly thinned. Selling an open-ended spending commitment with no backlog behind it isn’t a reflex, it’s underwriting. Two of those three are worth taking on their own, because they are where the divergence runs deepest and most durable: whether each company can actually pay for the build, and whether it will show you enough to check that it is working.The real divide is cashTake the funding first, because it is the starkest of the three and it barely made the coverage. Microsoft generated $19.6B of free cash flow in the quarter, on $41B of capex, and still returned $10.2B to shareholders. For the full year it threw off $182.9B in operating cash flow, funded a capex program that nearly doubled to $116B, returned more than $43B, and ended the year in a net cash position. It is paying for the largest infrastructure build in its history out of pocket and handing money back at the same time.Meta generated $784M of free cash flow. That is not a typo. Free cash flow fell 91% year over year, the company issued a fresh $24.9B of bonds in the quarter, and it suspended its buyback entirely, zero repurchases in the first half against a $26B pace last year. Full-year free cash flow is now tracking toward breakeven. Meanwhile stock comp accelerated to $13.7B in the first half, up 52%, so the share count now faces real dilution with no buyback to offset it.Same capex race. One company is self-funding it with room to spare. The other is borrowing to fund it and has stopped buying back stock to preserve the cash. And Meta isn’t even the extreme case: Alphabet, reporting the week before, posted its first negative free cash flow quarter since 2004 while raising its capex guide for the third time this year.This is the real 2026 dividing line among the hyperscalers, and it isn’t the one the who’s winning AI framing captures. Everyone is spending. Everyone has demand that, by their own telling, exceeds supply. What differs is whether the spending comes out of a cash flow statement that can absorb it or a balance sheet that has to stretch. Microsoft can be wrong about AI demand and the cost is a few years of lower returns on capital. Meta, borrowing into a near-zero free cash flow year with the buyback already switched off, has less room for the same mistake. The market is right to price that difference, and it is more durable than any single quarter’s revenue growth.Both companies got asked the question that matters, and both dodged it.At Microsoft, Goldman’s Gabriela Borges asked directly for a quantified return on the capex, how the ROIC math compares to a year ago. CFO Amy Hood’s answer was entirely qualitative: “I don’t know that my math has changed... confidence in the TAM expansion, the margin levers.” A pointed financial question, a narrative answer, on a quarter with $41B of capex. At Meta, Morgan Stanley’s Brian Nowak asked for early color on 2027 capex. CFO Susan Li declined outright, calling planning “highly dynamic.” No number, on the single biggest unknown in anyone’s model.Two management teams, same evasive posture on the one input the whole valuation turns on. But the market treated them completely differently, and the reason is disclosure. Microsoft could deflect the ROIC question because it handed investors something concrete to hold instead: commercial bookings excluding OpenAI more than doubled their growth rate, from 7% to 18%.That excluding OpenAI is worth pausing on, because it is the same footnote game running underneath the headline EPS. Microsoft’s backlog is so distorted by the OpenAI commitment that the company has to strip it out to show a clean demand number: remaining performance obligations grew 84% including OpenAI and 25% without it. So the very stake that flatters Microsoft’s GAAP earnings on the income statement is the one it removes from its bookings to prove the core business is accelerating. It gets OpenAI both ways, an earnings tailwind you leave in and a distortion you take out, depending on which number you want to look good. The saving grace is that the clean signal genuinely is clean: that $678B backlog grew while its average duration compressed to 2.3 years from 2.5, and stretching duration is precisely how a backlog gets flattered. Microsoft’s got bigger while getting shorter, the opposite of window dressing. Copilot paid seats went from 20 million to over 30 million. There is a scoreboard.Meta had no equivalent. Its strongest defense of the buildout was Li’s line that the business is “demand-constrained” with “ROI-positive places we would put compute if we had it,” which is a claim, not a metric. There is no backlog figure, no bookings inflection, nothing an investor can mark against the $130-145B it plans to spend this year. When you spend like a hyperscaler but disclose like a consumer-ads company, the market makes you prove it every ninety days. That is a big part of the gap between plus-14 and minus-9. It isn’t that Meta’s business is worse. It’s that Meta gave the market less to believe.Now read the footnotesHere is where the two stories converge rather than diverge, and where both get a little worse. Under the surface, each is telling you less than it did a year ago, in two ways at once: how the build is financed, and how it’s measured.Start with the financing. In fiscal 2027 Microsoft is extending the useful life of its data centers and reclassifying future data-center leases from finance leases, which count in capex, to operating leases, which don’t. Hood said it plainly: “outside of this useful life impact, our calendar year 2026 CapEx investment expectations remain unchanged,” but the reported figure “adjusts to approximately $175 billion.” The reported capex line is about to shrink for accounting reasons while the real spend doesn’t. Meta is doing a version of the same thing from the other side of the balance sheet: it funded this quarter partly with $24.9B of fresh bonds and announced a joint venture with BlackRock to build a one-gigawatt data center in El Paso. Between the debt and the JV, a growing share of Meta’s build is moving off its own balance sheet into structures that don’t read as cleanly as capex paid from cash. The pattern across both is the one now dominating the AI credit conversation: as the numbers get bigger, more of the spending is being routed through the balance sheet’s back rooms, leases that don’t count as capex at Microsoft, bonds and partnerships at Meta, exactly when investors most need a clean read on how much is really being spent and who is on the hook for it. The number to trust at Microsoft next year is the cash flow statement, cash paid for property and equipment plus finance-lease additions, not the headline capex line management will point to.Meta’s capex guide narrowed to $130-145B this quarter with no stated rationale, where a quarter earlier it had explicitly blamed a raise on memory-price inflation. And both companies retired their cleanest AI-monetization metric in the same quarter. Microsoft dropped the “$37 billion AI ARR, up 123%” figure Nadella leaned on last quarter, replacing it with a vaguer “Foundry customers at 1 trillion tokens, up 4x,” which is activity, not revenue. Meta’s broadest AI-ads number, Advantage+ at a $75B run rate, doesn’t reconcile cleanly with the narrower $20B “Value Optimization Suite” figure it cited a quarter earlier, so you can’t actually track the trend.None of this is fraud, and none of it breaks either thesis. But the pattern is worth naming. As the capex climbs, both the financing and the metrics are drifting toward places investors can’t see clearly, at exactly the moment the stakes say they should be getting sharper. Microsoft can afford the opacity because it backed this quarter with a real bookings acceleration and self-funds the build. Meta is asking for the same faith with a weaker balance sheet and no backlog to show. Same move, very different amounts of credibility to spend on it.What it meansThe larger point is about how to read the rest of this cycle. The market has started underwriting AI capex the way a lender underwrites a project, demanding the two things that separate an investment-grade build from a speculative one: that the borrower can fund it from its own cash, and that a contracted backlog shows the capacity is already spoken for. Microsoft presented investment-grade on both counts and got a re-rate. Meta presented closer to speculative, borrowing into the build with no backlog to point to, and got punished, even with a sound quarter underneath. In a season when good prints have mostly been sold, this pair is the exception that explains the rule: the beat with the self-funded story and the backlog got bought, the beat without got sold. The beat itself was never the variable.Both are longs in my opinion, but they are not the same kind of long, and the last two sessions widened the gap. Microsoft is the higher-quality, lower-risk name. Self-funded, net cash, a $678B backlog, and a print that just retired the one fear hanging over it, so that even an unkind three years, margins giving a little back and the multiple parked at the bottom of its range, still leaves you with a small gain rather than a loss. That is about as close to a floor as a mega-cap at this price offers. The catch is you now have to pay for it. The stock is up roughly 14% in two sessions, to about $447 from $391, and the easy money, buying the margin-fear washout into the print, was there for anyone who owned it Tuesday afternoon. The required-return math still clears a double-digit annual return on the base case, but the margin of safety is thinner than it was 48 hours ago and the catalyst has already fired. Microsoft from here is a compounder to hold, not a pitch to swing at.Meta is the pitch to swing at, and it is the more attractive risk/reward at these levels precisely because it got hit. The market was right to mark it down and the forward worries are real, but the drop to about $532 took the stock below where it traded in the spring tariff panic while the value of the business underneath didn’t move with it. The downside from here looks bounded: the bear case, where opex never decelerates and the multiple de-rates toward ad-tech levels, prices into the high-$490s, less than 10% below today. Against that, a business still compounding revenue in the mid-20s at high-30s core margins is worth roughly $640-720 on a fair near-term multiple and well into the $900s over two-to-three years if the forward cracks prove cyclical. Bounded single-digit downside against 20% near-term and far more over the cycle is a lopsided payoff, and you are being paid it to underwrite a question the company answers on a fixed schedule rather than an open-ended one.The honest catch on Meta is that its bear case isn’t a stub. If R&D keeps growing at twice revenue and impressions keep sliding, the multiple compresses further and cheap gets cheaper, the textbook value trap. This is the higher-variance long, and it earns its place for an investor who can sit through a near-zero free cash flow year and a couple of noisy prints. Microsoft is the one to own for the compounding without the drama. Meta is the one that pays you for tolerating the drama, and right now it pays more.What to watch over the next three to six monthsThe whole debate resolves on a short list of dated, checkable signals, most of them landing in the same late-October print cluster.For Meta, it runs through two reports. Q3, in late October, is the first read on the swing variable: R&D growth ex-severance decelerating toward the 40s% confirms the cost bulge is temporary and hands the bulls their argument, while holding near 55-60% validates the bear. Watch ad impressions in the same print, another leg below +14% would turn the revenue-quality crack into a trend rather than a quarter. Then the Q4 print, likely late January, is when Meta has historically set next-year capex, so it is the moment the company either finally bounds its 2027 spending or extends the open-ended commitment that is half of why it is being discounted. A buyback resumption anywhere in that window would be management signaling its own cash worry has passed. Sooner than any of it, watch whether the sell-side re-anchors to the ex-items earnings number over the next couple of weeks, the mechanical retrace the headline miss set up.For Microsoft, the one line that matters is Microsoft Cloud gross margin in the Q1 print, also late October. Holding at 65% buries the margin scare for good; resuming its slide while capex keeps climbing brings it right back and takes the re-rate with it. Alongside it, check that Azure actually delivers the ~45% it just guided and that bookings ex-OpenAI hold near 18% rather than reverting to 7%. The subtler thing to track all year is the capex line itself: the lease reclassification is about to make the reported number understate real spending, so the honest read is cash paid for property and equipment plus finance-lease additions in the cash flow statement, not the headline the company will cite. If that cash number keeps climbing while the reported line flatters, the margin-durability question the market just forgave comes back.And tonight, Amazon. The tell isn’t how fast AWS grew, it’s whether the capex guide goes up again and whether the cash flow can still carry it, the same two questions that split Microsoft and Meta. An Amazon that raises capex and stays firmly free-cash-flow positive reinforces exactly the divide this quarter drew: the market will keep paying up for the hyperscalers that can fund the build from cash, and keep making the ones that can’t prove it every ninety days. The spending is the same story across all of them. The balance sheets, and the willingness to open the books, are not. That’s the variable that sorts the winners from the merely busy.