Orders Are Easy. Profit Is Hard.

The Big Question

The four largest hyperscalers are on pace to spend roughly $730 billion on AI infrastructure in 2026. That number is real. Goldman Sachs projects total hyperscaler capital expenditure from 2025 through 2027 will reach $1.15 trillion, more than double the $477 billion spent in the preceding three years. JPMorgan raised its global AI capital expenditure estimate through 2030 to $5.5 trillion in June. The demand signal is not in dispute.

The question every serious portfolio manager is sitting with right now is a different one: of all that capital flowing into AI infrastructure, how much actually becomes profit for the companies assembling the hardware? And if the answer is “very little,” which companies are structurally positioned to be the exception?

Tonight, after the market closes, Supermicro reports its Q4 fiscal 2026 results. The number on every institutional desk is not the revenue figure. It is the gross margin. And behind that number sits one of the more important debates in technology investing right now: whether the AI infrastructure buildout is creating durable, high-quality earnings for the hardware companies executing it, or simply funding a very expensive land grab where the spoils go to someone else.

Why Wall Street Cares

For most of 2024 and into 2025, the AI infrastructure trade felt straightforward. GPU demand drove Nvidia. Nvidia’s allocations drove server OEM volume. Volume drove revenue. The market paid for growth, and the growth was real.

2026 broke that logic. The 30-day correlation between major AI spenders and semiconductor stocks collapsed from 0.78 to near zero, the lowest reading in four and a half years. Alphabet and Amazon rose on strong cloud revenue that gave investors confidence their infrastructure spending was converting into returns. Meta and Microsoft initially declined when they announced more aggressive spending, because investors focused on the scale of outlay relative to near-term revenue visibility. The market started asking a harder question: who in the AI supply chain actually earns the economics of this boom?

That question lands hardest on server OEMs. They sit between the GPU suppliers who set chip prices and the hyperscalers who set server prices. When both sides squeeze, margin disappears. Supermicro spent the better part of fiscal 2026 with gross margins in the 8-to-10% range, operating in a business that once earned 14% to 17% before the AI volume surge commoditized the product mix. The preliminary numbers for Q4 suggest something changed. Whether that change is structural or coincidental is what institutional investors are debating today.

The Bull Case

The bull argument starts with a remarkable fact. Supermicro’s Q4 preliminary update guided gross margins to 15% to 17%, against prior guidance of 8.2% to 8.4%. That is not a rounding error. It is a 650-to-700-basis-point swing driven, management said, by a favorable customer and product mix. In plain terms: high-density, liquid-cooled AI server configurations for specific, sophisticated buyers carry meaningfully different economics than commodity server volume.

The deeper bull argument is about the nature of Supermicro’s moat. The company built the original xAI Colossus cluster in 122 days. CEO Charles Liang posted in June 2026 that Supermicro was co-building another gigawatt-scale AI data center for SpaceX and xAI, calling it the fastest build time yet. For frontier AI labs and hyperscalers, deployment speed is not a soft preference. Every day a cluster sits unbuilt is a day of GPU capacity that cannot generate revenue. That urgency is worth paying for, and it is why $60 billion in new orders arrived in a single quarter.

On valuation, the bull case is equally pointed. Supermicro’s forward non-GAAP price-to-earnings sits around 11x, versus an industry average near 24x. The company has nearly doubled revenue year-over-year. Needham raised its price target to $46 after the preliminary update, citing the margin step-up as evidence that the business mix is genuinely shifting toward higher-value work. At 18x to 20x forward earnings on $2.97 in FY2027 EPS, the stock has a straightforward path above $50.

The Bear Case

The bear argument does not dispute the demand. It disputes whether Supermicro can capture it as profit rather than as working-capital consumption.

The numbers from Q3 fiscal 2026 make the concern concrete. Operating cash flow was negative $6.6 billion in a single quarter. At the same time, the balance sheet showed $8.8 billion in bank debt and convertible notes alongside $11.1 billion in inventory and $8.4 billion in receivables. The company had to raise $7 billion in equity and equity-linked financing in June just to fund component purchases for the backlog. Reported earnings were growing. Cash was disappearing. The two things were happening simultaneously, and that is not a temporary artifact.

Then there is the legal dimension. In March 2026, the DOJ unsealed an indictment charging three individuals tied to Supermicro, including co-founder Yih-Shyan “Wally” Liaw, with conspiring to illegally export over $510 million in Nvidia-equipped AI servers to China. Supermicro is not named as a defendant. But the reputational and operational damage extends beyond the individuals charged. A class action lawsuit filed by shareholders in California alleges the company failed to disclose export-control violations. An independent board investigation, supported by Munger, Tolles and Olson and forensic accountants from AlixPartners, has no definitive timetable for completion.

Goldman Sachs, holding a Sell rating, has cited ongoing margin pressure and limited visibility into profitability in Tier 2 cloud markets as its core concern. The stock has lagged its two closest AI-server peers badly in 2026, up roughly 9% year-to-date against Dell’s 242% gain and HPE’s 103% advance, all three riding the same underlying infrastructure wave.

The Evidence

The data defines the tension clearly.

• Q3 FY2026 revenue: $10.24 billion, up 123% year-over-year
• Q3 non-GAAP EPS: $0.84 (vs. $0.31 in Q3 FY2025)
• Q3 gross margin: 10.1% non-GAAP, up from 6.4% the prior quarter
• Q4 preliminary gross margin: 15% to 17%, versus 8.2% to 8.4% prior guidance
• Q4 new orders: over $60 billion, a company record
• Q3 operating cash flow: negative $6.6 billion
• June 2026 equity raise: $7.0 billion in equity and equity-linked financing
• Wall Street Q4 consensus EPS: $0.92 to $0.96; revenue consensus near $11.56 billion
• FY2027 EPS consensus: $2.97, implying 41% growth over FY2026
• 52-week range: $19.48 to $58.78

The competitive context sharpens the evidence further. According to IDC data, Supermicro leapfrogged both Lenovo and HPE to become the second-largest server vendor globally by the end of 2025, behind only Dell. Its deployment speed advantage is documented: lead times of two to four weeks versus ten or more weeks for HPE during constrained periods. Yet Dell, which now targets $60 billion in AI server revenue for FY2027, is winning the stock market’s confidence where Supermicro is not. Dell trades at roughly 34x trailing earnings. Supermicro trades at 17x. Both are executing on the same AI infrastructure cycle.

That gap has a source. The preliminary update itself noted that some of the $60 billion in orders may not constitute binding purchase commitments and could be subject to cancellation or delay. The SpaceX and xAI relationship, while confirmed as active by Liang’s public statements, remains unquantified. On the original Colossus build, Supermicro was a co-supplier: Dell assembled roughly half the racks on its PowerEdge platform while Supermicro provided the other half. The assumption that one customer relationship explains the bulk of the $60 billion backlog is a market inference, not a disclosed fact.

The Mavens’ View

The institutional debate is not about whether Supermicro has demand. It is about whether this company can convert demand into durable, high-quality cash earnings without continuously diluting shareholders or consuming its own working capital to fund growth.

Hedge fund positioning offers a signal. According to Insider Monkey’s database, 49 elite hedge funds held stakes in Supermicro at the end of Q1 2026, up from 39 the prior quarter. That is not the positioning of a market that has written the company off. Marshall Wace, one of the more data-driven multi-strategy funds, held a position worth $163.6 million. The accumulation was happening even as the DOJ headlines were breaking.

What sophisticated investors appear to believe is this: the legal overhang is real but company-specific rather than existential. The three individuals charged are no longer at Supermicro. The company is not named as a defendant. The independent board investigation, while carrying no timetable, is being run by credible outside counsel and forensic accountants. That is a different situation from 2023, when Supermicro faced accounting restatements and a short-seller report that cast doubt on the fundamental business.

But the margin question is genuinely unresolved. Mizuho’s Vijay Rakesh, who carries an 84% accuracy rate among tracked analysts, cut his price target from $44 to $34 after the preliminary update, citing the revenue shortfall and ongoing supply constraints. That is not a bearish call on the margin improvement. It is a skeptical call on whether the margin improvement was driven by repeatable product mix or one specific customer’s configuration requirements in a single quarter. The two things look identical in the preliminary data and completely different in what they imply for the next twelve months.

Barclays sits at $38. Needham sits at $46. The gap between those two targets reflects the unresolved question exactly. The answer arrives tonight.

What Investors Are Missing

Almost nobody is discussing the second-order consequence of Supermicro’s legal exposure on its competitive positioning, specifically on GPU allocations.

Nvidia publicly distanced itself from the DOJ indictment, emphasizing that its chips include export compliance features. But the practical implication of a major server OEM facing ongoing export-control scrutiny is that Nvidia has both the incentive and the commercial cover to quietly redistribute GPU allocations toward cleaner counterparties. Dell and HPE unveiled Vera Rubin AI server platforms at Nvidia GTC 2026 at almost the exact moment the indictment news broke. Futurum Research called it directly: the indictment creates a GPU allocation opening for Dell and HPE. Dell has a stated AI server revenue target of $60 billion for FY2027. If any portion of Supermicro’s allocation has quietly shifted, the $60 billion backlog may face a component constraint that has nothing to do with memory shortages or shell construction delays.

The second overlooked implication is the cash conversion cycle as a forward indicator. Morgan Stanley estimates that nearly $3 trillion in AI-related infrastructure investment will flow through the global economy by 2028. If Supermicro is burning $6.6 billion in operating cash to build $10 billion in quarterly revenue, the working-capital intensity of that growth trajectory becomes a structural constraint on how much of the next $3 trillion the company can actually capture. The $7 billion equity raise was not a one-time event. It was a signal about the business model’s capital requirements at scale. The question investors should be asking tonight is not whether margins improved. It is whether cash generation improves alongside margins, or whether one simply substitutes for the other in the quarterly summary.

Stocks to Watch

Super Micro Computer (SMCI). The central name in the debate and the most binary outcome tonight. If Q4 confirms the 15-to-17% gross margin, Q1 FY2027 guidance comes in above $12 billion, and management provides any credible explanation for how operating cash generation improves as backlog converts, the rerating case is straightforward. At 11x forward earnings with the margin story intact, the discount to peers is hard to justify. If management is vague on backlog composition, if cash flow guidance is absent, or if Q1 gross margin guidance reverts toward 12% or below, the discount is earned and the legal overhang amplifies it. This is a two-scenario stock in a way few others are tonight.

Dell Technologies (DELL). The cleaner version of the same trade. Dell has been the dominant AI server stock of 2026, up 242% year-to-date through late June. It booked $24.4 billion in AI orders in Q1 FY2027 alone and targets roughly $60 billion in AI server revenue for the full fiscal year. It trades at a premium to Supermicro for a reason: no governance overhang, stronger balance sheet, broader enterprise relationships. If Supermicro’s legal exposure continues to suppress GPU allocations from Nvidia, Dell is the primary beneficiary. It is also the floor-case investment: if AI infrastructure demand proves durable but SMCI stays impaired, Dell captures the displaced share.

Hewlett Packard Enterprise (HPE). The overlooked name in the AI server conversation. HPE’s stock is up 103% year-to-date, well ahead of SMCI, and it unveiled the Vera Rubin NVL72, a fully integrated rack-scale liquid-cooled system supporting 72 Nvidia Rubin GPUs, at GTC 2026, with availability targeted for December 2026. Its heritage in high-performance computing through the Cray acquisition gives it credibility with the sovereign AI programs and national laboratory customers that are increasingly writing large checks. If Supermicro’s compliance scrutiny makes procurement officers at government-adjacent buyers nervous, HPE is the natural redirect.

Nvidia (NVDA). The most important company in this entire debate, and the one with the most pricing power at every step. Nvidia controls GPU allocations. It sets the component cost structure that every server OEM lives within. Its data center revenue reached approximately $93.7 billion in fiscal year 2026. Regardless of which server OEM wins the margin argument, Nvidia collects its economics on every GPU that ships. Tonight’s Supermicro call will offer a real-time read on the health of the AI infrastructure order cycle from the integrator’s perspective. If Supermicro’s $60 billion backlog is real and converting, that is additional confirmation of the Nvidia demand cycle that the market is already pricing at a premium multiple.

Arista Networks (ANET). The quietest winner in this debate. Every gigawatt-scale AI cluster that Supermicro, Dell, or HPE assembles requires high-speed networking infrastructure. Arista supplies spine-and-leaf Ethernet fabric to hyperscalers and frontier AI labs. It carries none of the governance risk of the server OEMs, has demonstrated consistent gross margins above 60%, and benefits from the same $730 billion infrastructure spending wave without sitting between GPU suppliers and cost-sensitive hyperscaler procurement teams. If the debate about AI hardware profitability keeps the market cautious on server OEMs, Arista is where institutional investors looking for clean AI infrastructure exposure are increasingly parking capital.

Bottom Line

The AI infrastructure spending cycle is not in question. What is in question is whether Supermicro has crossed a structural threshold in its ability to profit from it, or whether Q4 was a favorable quarter of customer mix that will mean-revert as the $60 billion backlog gets fulfilled across a broader product set at more ordinary margins.

The investment committee answer to watch for tonight is not the headline EPS. It is the Q1 FY2027 gross margin guide, the operating cash flow direction, and anything Liang says about customer concentration within that backlog. A company that can book $60 billion in orders in ninety days and convert them at 15% gross margins with improving cash generation is one of the most mispriced AI infrastructure stocks in the market. A company that does the first part but not the second is exactly as cheap as it deserves to be.

Those are two completely different investments wearing the same ticker. Tonight’s call is the test.