AMD Guided 5.3% Above Consensus. So Why Is Money Still Leaving?

The headline results:

"non-GAAP" is the company's adjusted version that strips out certain costs. "Guidance" is management's own forecast; "consensus" is the average forecast of Wall Street analysts.)

Line up the last five quarters and you see revenue climbing more steeply while gross margin moves sideways:

In one year, Data Center went from less than half of revenue to nearly two-thirds. Normally that would pull overall margins up. Instead, gross margin has bounced between 55% and 57%. Why? The most plausible explanation is that the new generation of Instinct AI chips and the Helios full-rack system are still early in their production ramp. Manufacturing yields, packaging and high-bandwidth memory (HBM) costs haven't been spread over enough volume yet, so lower GPU margins are temporarily offsetting the high margins of AMD's EPYC server processors. AMD doesn't break out GPU margins separately, so this explanation can only be tested by watching where margins go over the next few quarters.

AMD doesn't give full-year guidance, but adding the Q4 forecast makes the math simple. Revenue for Q1 through Q3 totals $35.17B; add the $17B Q4 midpoint and FY2026 comes to about $52.2B, roughly 50.6% above FY2025's $34.64B. Q4 alone would need to grow 27% over Q3. On the call, management said Data Center should grow about 50% sequentially in Q4 (as paraphrased in call transcripts, 2026-11-03). Off an $8.5B base, that's about $12.75B. That implies Client & Gaming plus Embedded only need to bring in about $4.25B combined in Q4, more than 10% below their $4.8B in Q3. So the one number that really matters in the next report is whether Q4 Data Center revenue lands near $12.75B. Even if total revenue hits the target, if it gets there on the back of PCs and embedded chips, that would mean Helios is ramping slower than promised, and the most expensive part of the valuation would lose its footing.

The evidence on demand is firmer than last quarter. In management's words: "Demand for our Instinct and EPYC products continued to strengthen, and we expect growth to accelerate further in the fourth quarter as we scale Helios and Venice production." (company 8-K, 2026-11-03). Helios is AMD's full-rack AI system; Venice is its next-generation EPYC server CPU. The 8-K names Data Center customers including Anthropic, Meta, Microsoft, OpenAI and Oracle Cloud Infrastructure, with HP and Lenovo expanding their system offerings. Anthropic alone accounts for about 2GW of compute power capacity (company 8-K, 2026-11-03), and media reports say the first 1GW is expected to come online in the first quarter of 2027 (itiger and others, June 2026 reports). MI450 and Helios have already started contributing revenue this quarter (earnings call, 2026-11-03).

AMD is a profitable "fabless" chip company: it designs chips but outsources wafer manufacturing and packaging to TSMC. It is not burning cash to grow. So the question for this section is the quality of its profits: where the GAAP and adjusted figures differ, how much profit actually turns into cash, and how much existing shareholders will give up under a warrant deal signed long ago.

Start with the accounting gap. This quarter's non-GAAP EPS of $2.09 versus GAAP EPS of $1.46 is a $0.63-per-share difference. Working backward from GAAP net income of $2.43B and GAAP EPS gives about 1.66 billion diluted shares, so the two measures differ by about $1.05B for the quarter. Based on the company's usual reconciliation, most of this is amortization of intangible assets from the Xilinx and ZT Systems acquisitions, plus stock-based compensation (SBC). Amortization is a bookkeeping cost of past deals and no longer uses cash. SBC, however, is a real cost to shareholders. The exact SBC amount for this quarter will come in the 10-Q filing; the earnings release summary doesn't list it separately.

Cash conversion is the most glaring issue this quarter:

Think of a restaurant that just had record sales but has the thinnest cash drawer of the year, because the money is tied up in ingredients stocked in the kitchen and big tables that haven't paid their bills yet. For AMD, the "ingredients" are inventory built up for the Helios ramp, and the "unpaid tables" are receivables from big customers. The reason this dip looks acceptable is that it points the same way as the 27% sequential growth in Q4 guidance: you stock up before you ship. And a 77% year-to-date conversion rate shows earnings quality hasn't collapsed. But if Q4 delivers $17B in revenue and free cash flow doesn't return to roughly last year's $1.5B level, it would be fair to suspect that big customers are stretching out their payment terms.

Dilution is the other account to settle. In October 2025, when AMD signed a 6GW Instinct deployment agreement with OpenAI, it also granted OpenAI warrants for up to 160 million shares at an exercise price of $0.01 per share. They vest in tranches as deployments scale from the first 1GW to 6GW, and each tranche carries a share-price hurdle; the final hurdle is about 23% above the current share price (company announcement, 2025-10-06). Based on roughly 1.63 billion shares outstanding (market cap of $797.9B divided by the latest share price, 2026-11-04), full vesting would dilute existing shareholders by about 9.8%. This dilution buys a 6GW order anchor, and both sides of that trade matter. Because the warrants are tied to deployment, dilution happens exactly when revenue is delivered, so in effect AMD is trading shares for certainty. But with an exercise price of practically zero, once vested it amounts to nearly a tenth of the company in free shares, and the related GAAP expense will widen the gap between the two earnings measures further.

There are also reports in the market that Anthropic has committed to an equity stake of about $5 billion in AMD (media reports, mid-2026). We could not verify this in the company's own filings, so we are not treating it as a source of funds that has actually happened. AMD doesn't need outside financing right now: $4.68B of year-to-date free cash flow comfortably covers R&D and capital spending. Its funding risks lie mainly in dilution and receivables, not in debt repayment.

Upstream, there are three hard bottlenecks, each more expensive than the last. TSMC is AMD's main source for both wafers and advanced packaging. CoWoS-L (an advanced packaging method that puts multiple chips and HBM on a single base) is expected to remain the dominant approach through 2028, with capacity staying tight (TrendForce, 2026-09-18). HBM (high-bandwidth memory stacked next to the GPU) remains supply-constrained, with its blended average selling price expected to rise 121% year over year in 2027 (TrendForce, 2026-09-29). Server demand driven by AI inference is also expected to keep high-capacitance MLCCs (small passive components on circuit boards) in short supply into the first half of 2027 (TrendForce, 2026-10-29). On the rack side, after acquiring ZT Systems in 2025, AMD sold its manufacturing operations to Sanmina, and Helios rack assembly relies on that outsourced line (company announcement, 2025). For a chip designer, rising upstream prices mean rising costs. Products like the MI450 carry a lot of HBM per card, so HBM makes up a very large share of their cost, which is another reason gross margin has been stuck around 56%.

Downstream, the buyers are highly concentrated. Data Center is about 64% of total revenue, and most of that comes from a handful of cloud providers and AI labs:

These contracts need to be read carefully. A gigawatt (GW) is a unit of data center power capacity. Signing for 2GW means power has been reserved for computing expansion; between that and actual chip shipments and booked revenue lie three hurdles: building the facility, delivering the hardware, and customer acceptance. "Up to 6GW" and "about 2GW" are ceilings. The only thing that actually lands is the revenue recognized each quarter.

The weakest link in the chain is actually the customer's customer. OpenAI and Anthropic are still losing large amounts of money and funding their growth by raising capital, so their gigawatt commitments ultimately depend on their ability to keep raising money. If private-market enthusiasm for AI lab valuations cools, the first things to get pushed back will be these not-yet-built gigawatts, and AMD's two largest sources of new growth both fall into this category.

All six computing-chip peers have reported this quarter. Here's what can be compared (peer disclosures, as of 2026-11-04):

Side by side, AMD's position is clear: its total revenue growth is less than half of NVDA's, its gross margin is 20 percentage points lower, and the power capacity it has locked in is a quarter of NVDA's. But AMD's 45% is dragged down by the -8% in Client & Gaming. Data Center alone grew 97%, in the same league as NVDA's overall 106%. The real gap is in gross margin: that 20-point difference is the price tag on NVIDIA's CUDA software moat, and it caps how much profit AMD keeps on each extra card it sells. On demand, AMD, MRVL and AVGO all reported strengthening demand, all pointing the same way. This upswing is still moving the entire computing supply chain together. AMD hasn't fallen behind; it's simply in the second tier.

How does the market price report cards like these? Three semiconductor peers that have already reported gave a fairly chilly answer:

(The "implied move" is how much the options market expects a stock to swing after earnings.)

INTC's plunge is the most telling: a huge EPS beat was met with a drop far beyond what was implied, showing the market no longer pays for a "beat" itself, only for whether guidance and the path to delivering it can keep rising. INTC is also the peer most closely tied to AMD. Over the past 8 quarters, the two stocks moved in the same direction after earnings 6 times, with a correlation of 0.78; for TXN it was 5/8 and 0.41, and for MU 3/8 and -0.35 (data as of 2026-11-03). But 8 quarters is a short sample, so this relationship is only a reference for how sentiment spreads, not a rule. AMD's own implied move this time was ±6.4%. Over the past 8 quarters its actual move exceeded the implied move 4 times, with a median implied move of ±7.0% against a median actual move of 6.7%, so options have priced AMD's earnings fairly on the whole. On the first trading day after the report, AMD rose 1.0% (2026-11-04), well inside the implied range. That suggests most of the good news was already expected: the market didn't punish AMD the way it punished INTC, but it didn't give it an extra reward either.

A trailing P/E of 109 has lost its usefulness as an anchor. Earnings grew 178% in a year, so the denominator is changing too fast, and GAAP figures are depressed by acquisition amortization. Using it to judge whether the stock is cheap or expensive will only ever tell you "too expensive." Better yardsticks are a forward P/E based on annualized non-GAAP earnings from the latest quarter, EV/Sales based on annualized Q4 guided revenue, and PEG to adjust for growth.

(P/E is share price divided by earnings per share; EV/Sales compares the company's total value to its revenue; PEG divides P/E by the earnings growth rate.)

What does 57 times mean in practice? If you set the reasonable end point at the 30x forward P/E that's common for a mature, high-growth semiconductor company, today's market cap would need about $26.6B in annualized non-GAAP net income to support it. That's about 1.9 times the Q3 annualized level and 1.5 times the Q4 guided annualized level. In other words, the market has already paid in advance for Data Center to roughly double again in 2027, which means the gigawatt-scale orders from OpenAI, Anthropic and Oracle need to land more or less on schedule. A PEG of 0.61 looks cheap, but only because it uses 178% growth, part of which comes from last year's low base. Drop the growth rate to 40% and the PEG at 57 times rises to about 1.4: not cheap, but not outrageous either. Room for the multiple to expand depends on earnings continuing to grow at a doubling pace. If growth falls back to the industry average, the multiple will inevitably compress.

The AI computing cycle has its own rhythm: cloud providers' capital spending (capex) drives demand; upstream wafers, advanced packaging and HBM determine supply; and chip designers sit in the middle, profiting from the spread between the two. Three types of readings help locate where we are:

Our view: this computing cycle is in the middle-to-late part of its upswing. Demand is still accelerating, and every upstream link is raising prices and running short, which is classic mid-cycle behavior. But upstream price hikes have moved from "expected" to hard quotes of a doubling in 2027 ASPs, so chip designers are starting to feel cost pressure. Downstream customers are locking in power and capacity for 2027, and every link in the chain is stockpiling ahead of time, which are signs the cycle is moving toward its later stage. For AMD, mid-to-late cycle means revenue keeps rising while gross margin gets squeezed by upstream price increases, and the spread starts to narrow.

Two trackable early signs of a peak: first, quarterly increases in HBM and DRAM contract prices start to narrow noticeably; upstream prices that stop accelerating often lead the demand peak by one to two quarters. Second, in cloud providers' fourth-quarter reports from late January to early February 2027, 2027 capex guidance growth steps down significantly from this year's doubling pace. If both show up together, the cycle call should move from mid-to-late stage to the peak zone.

This section deals with weekly signals that govern entry and exit timing over the next few weeks. Readers holding for a year or more can skip it.

Money is leaving AI chips, and leaving fast. As of 2026-11-04, the AI chip sub-sector that AMD belongs to has dropped into the weakest tier of the whole market for 3 straight days, its relative strength sits in the lagging quadrant, its sector money flow reading is -50, and it has seen net outflows for 10 consecutive days. Net flow over the past 5 days totals -5.8B shares, while the past 20 days total only -955.1M shares (both estimated from price and volume as shares traded, not dollars). A 5-day outflow bigger than the 20-day outflow means the two-plus weeks before were actually net inflows, and the selling burst out in the past week, right after INTC's post-earnings plunge. The sub-sector's price fell 18.8% in 5 days, faster than the money left. That's what panic selling looks like; an orderly exit by long-term investors doesn't produce a slope that steep. The sub-sector was downgraded from the neutral tier on 2026-11-02, and its parent sector, Technology, is also in the weakest tier.

Where did the money go? On the same day (2026-11-04), the top three sub-sectors for inflows were storage REITs (PSA, reading 77), exchanges and financial data (CME, 76) and IT services (ACN, 74). The top three for outflows were lithium batteries (LIT, -68), the semiconductor index (SMH, -58) and quantum computing (IONQ, -58). The two strongest rotation pairs both ran from Technology (XLK) into Real Estate (XLRE, strength 4.56) and Materials (XLB, strength 4.49). The broad rotation signal is risk_off. Consumer discretionary, industrials and healthcare each have their own rotation threads, but AMD's sector isn't on any rotation path right now: no fresh money is waiting to pick up the baton.

The macro backdrop isn't helping either. Net liquidity stands at $5.75T, down 1.0% week over week and contracting, and the 10-year Treasury yield is 4.95% (FRED, 2026-11-04). The broad market is neutral: SPY at 766.81 versus its 20-day average of 767.59, a gap of just -0.10%. The index isn't weak; tech and semiconductors are. This is a sector-level retreat, not a systemic risk.

History doesn't point either way. In the 20 trading days after falling into the weakest tier, the median return is +0.0%, with a 55% chance of a gain (3080 instances, 2021-04-12 to 2026-10-06). That's close to a coin flip: the weakest tier doesn't predict further declines, it just means there's no statistical edge in the short term. AMD's own stock price is also in no-man's-land: about 11% below its 50-day moving average and about 18% above its 200-day moving average (2026-11-04). The medium-term trend is intact; the short-term trend has broken. The 1.0% gain the day after earnings came amid outflows across the sector, which counts as relative strength, but it's a long way from a reversal.

Long term (quarters to a year): worth following. Data Center growth of 97% year over year, Q4 guidance 5.3% above consensus with even the low end above consensus, and gigawatt-scale customers like OpenAI and Anthropic starting to turn into revenue: these three things support a multi-year growth thesis. A 57x annualized forward P/E and a 0.61 PEG say the stock is expensive, but expensive with earnings to back it. The costs you have to accept should be spelled out too: gross margin is pinned near 56% by upstream HBM price hikes, 20 points below NVDA; quarterly free cash flow is only about 23% of net income; full vesting of the OpenAI warrants would mean about 9.8% dilution; and the two largest sources of new growth are themselves still funding their expansion by raising capital. How to hold it should follow the checkpoints: if Q4 Data Center revenue comes in near $12.75B, non-GAAP gross margin holds above 55%, and free cash flow returns to the $1.5B range, the long-term case gets stronger. If Data Center comes in clearly below $12B, or customers delay their gigawatt deployments, it's time to reassess.

Short term (days to weeks): don't chase. The sub-sector is in the weakest tier with ten straight days of net outflows, money is rotating from tech into real estate and materials, the broad signal is risk_off, net liquidity is contracting, and the stock is still about 11% below its 50-day moving average. The good news in the report was already priced in by options-implied volatility, and the stock rose only 1.0% the next day. In the short term, you want to see two things happen together: the sub-sector's net outflows stop and turn into sustained net inflows, and the stock climbs back above its 50-day moving average. Until then, a good earnings report isn't enough to support buying against the flow of money.

How those conditions are defined, and the price levels at which to act, are laid out in the action plan below.