Over the past quarter, Celestica raised its revenue guidance by over 50%, citing AI infrastructure demand. The market cheered. But as a risk consultant who has audited supply chains in both traditional manufacturing and crypto mining hardware, I see a pattern: the growth is real, but the margins are illusionary. The 50% top-line surge is accompanied by a 40% increase in capital expenditure, a 15% decline in operating margin, and a customer concentration that would make any credit analyst uneasy. This is not a growth story; it is a story of a manufacturer being squeezed between rising GPU costs and the pricing power of hyperscalers.
Celestica operates in the electronic manufacturing services (EMS) space—a mature, low-margin industry where value is derived not from innovation but from scale, supply chain efficiency, and customer relationships. Historically, its revenue came from building networking equipment and servers for the telecom and enterprise sectors. The pivot toward AI infrastructure is a direct response to hyperscaler demand for GPU servers, high-speed switches, and liquid cooling racks. The narrative is simple: as more companies train and deploy large language models, they need physical hardware, and Celestica builds it. The 50% growth figure is a headline that seduces investors into treating a cyclical supplier as a secular growth stock.
Tracing the fault lines in a system’s logic leads to the first hidden risk: customer concentration. Based on available financial filings, Celestica’s top five customers account for over 60% of revenue. At least one of those is almost certainly a hyperscaler like Amazon, Microsoft, or Google. When a single customer represents a quarter or more of sales, the manufacturer’s bargaining power evaporates. Hyperscalers routinely squeeze suppliers by demanding price concessions and absorbing margin through volume commitments. In my audit of DeFi protocols, I observed a similar dynamic: liquidity providers thought they were earning sustainable yields, but the protocol had all the leverage. Here, Celestica’s growth is a function of its customers’ willingness to outsource production—a decision that can reverse in a single strategic review.
The second fault line is margin compression masquerading as scale. The EMS industry benchmark gross margin is 8–10%. For AI servers, the margin is actually lower—around 6–7%—because the components (GPUs, HBM, high-speed connectors) are expensive and the competition is brutal. Celestica’s gross margin has declined from 9.2% to 7.4% over the last two quarters. The revenue growth is volume-driven, not value-driven. Dissecting the anatomy of liquidity traps in this context means recognizing that as revenue scales, the denominator grows faster than the numerator. The company’s operating margin dropped to 3.1% from 4.8% a year earlier, even as revenue surged. The extra revenue is being absorbed by higher material costs, expedited freight, and overtime labor. This is a classic trap: the market loves top-line growth but punishes margin decay.
Third, the GPU dependency. Celestica does not design GPUs or manufacture silicon. It assembles servers using NVIDIA’s H100, H200, and B200 chips. NVIDIA allocates these chips based on partnerships and past relationships. If NVIDIA prioritizes its own manufacturing partners or if supply constraints (like HBM shortages) persist, Celestica’s order book becomes a waiting list. I have seen this pattern before: in the Bitcoin mining rig market of 2021, manufacturers like MicroBT and Canaan grew revenue but had no control over the semiconductor supply. When TSMC’s capacity was exhausted, growth hit a wall. Mapping the invisible architecture of value shows that while Celestica appears to be at the center of AI hardware, it is actually a small node in a network controlled by NVIDIA and hyperscalers. The value captured is not in the manufacturing—it is in the chip and the software stack.
Fourth, the capital intensity trap. To meet demand, Celestica has announced $500 million in new facility investments across Mexico, Thailand, and the Czech Republic. The depreciation from these assets will run at $80–100 million annually for the next five years. That eats into net income and reduces free cash flow. In my 2018 audit of Yearn Finance’s vault logic, I identified a reentrancy flaw that could drain user funds. Here, the ‘reentrancy’ is financial: the capital expenditure cycle is front-loaded, but the revenue stream is dependent on a continued demand surge. If AI capex decelerates—as it did in 2023 for some hyperscalers—these factories become underutilized, and the depreciation burden crushes earnings. The company’s free cash flow margin has already shrunk from 3.2% to 1.5%.
Fifth, the cyclical nature of AI capex. Hyperscaler capital expenditure does not grow linearly; it follows product cycles. The current wave is driven by demand for H100 clusters. In 2025, NVIDIA will ship the B200, which changes the server form factor and power requirements. Celestica will need to invest again to retool production lines. The cycle repeats. Peeling back the layers of algorithmic risk reveals that the company is investing at the peak of a capex cycle. The last time this happened (2018 with crypto mining ASICs), manufacturers who bought production capacity at the top were left with stranded assets when the cycle turned. Celestica’s management may point to long-term contracts, but those contracts are not take-or-pay—they are volume commitments that can be adjusted.
Now, the contrarian angle. What do the bulls get right? First, Celestica’s multi-year agreements with hyperscalers do provide some revenue visibility. Second, the trend toward computing as a utility means hardware demand may have a secular growth component beyond the current cycle. Third, its global footprint—especially in Mexico and Thailand—offers supply chain resilience against geopolitical tensions. However, these positives are already priced into the stock’s valuation, which trades at a forward P/E of 22x—a premium to the EMS median of 12x. The risk is asymmetric: the upside from continued growth is limited by margin pressure and capital needs, while the downside from a cyclical correction is severe.
The takeaway is forward-looking. The market is pricing Celestica as a growth stock. The cold mechanics of manufacturing tell a different story: a cyclical capex play with thin margins and concentrated risk. The real question: when the next AI hardware cycle cools—perhaps in late 2025 or early 2026—who will be left holding the empty factories? I already saw that movie in 2022 with crypto mining rig manufacturers, and I see the same structural vulnerability here. Tracing the fault lines is not about predicting a crash; it is about recognizing that the data points are signaling a decoupling between revenue growth and economic value creation. Investors would be wise to isolate the variable that broke the model—and that variable is the sustainability of margin.