KEY READ-THROUGHS FROM CELESTICA’S Q3 2026 EARNINGS CALL AND INVESTOR DAY

Celestica’s call strengthens the outlook for AI infrastructure spending while providing more differentiated signals about where economic value will accrue. The most important developments are Ethernet’s expansion into scale-up networking, commercially meaningful alternatives to proprietary GPU systems, continued growth in pluggable optics, and rising requirements for liquid cooling and power infrastructure. The principal negative implications concern incumbent suppliers’ share of incremental spending and the cash required to finance extraordinary hardware growth. The $40 billion 2027 revenue outlook is evidence of substantial awarded manufacturing activity, not an equivalent increase in industry capital expenditure or a guarantee of revenue for every associated supplier. The highest-conviction conclusions favor companies with identifiable product exposure, while distinguishing established relationships from unallocated market opportunities. Country labels below indicate corporate domicile. Impact assessments describe fundamental exposure, not assumed share-price reactions.

AI SEMICONDUCTORS AND ACCELERATED COMPUTE

BROADCOM CAN MONETIZE BOTH CUSTOM ACCELERATORS AND THE NETWORKS CONNECTING COMPETING COMPUTE PLATFORMS (READ-THROUGH 1)

Affected company: Broadcom (AVGO: United States). Direction: strongly positive. Magnitude: potentially large at the consolidated-company level, with high confidence in the commercial transmission mechanism but no defensible incremental revenue estimate from this call alone.

Celestica identified Broadcom silicon in several distinct architectures. OpenAI’s Jalapeño accelerator was co-developed with Broadcom, its scale-up fabric uses Tomahawk 6 switches, and AMD’s Helios scale-up switch incorporates 2 Tomahawk 6 chips. Celestica also disclosed 10 awarded 1.6T programs, including 4 scale-up and 6 scale-out programs, and expects 1.6T revenue to exceed 800G revenue in 2027 while 800G continues growing.

The critical implication is that Broadcom’s opportunity is broader than custom accelerators replacing merchant GPUs. Broadcom can capture accelerator revenue where it participates in custom silicon and networking revenue where the compute platform belongs to another supplier. AMD’s success can therefore benefit Broadcom rather than simply redistribute spending away from it. Ethernet entering scale-up adds another addressable network layer beyond conventional scale-out fabrics.

This matters financially because AI semiconductor revenue already represented approximately 56% of Broadcom’s latest reported quarterly revenue: $16.7 billion out of $29.6 billion. Celestica’s disclosure is therefore relevant to a major earnings driver, not a peripheral product category. However, Broadcom had already disclosed exceptional AI growth and substantial forward demand. The new information is additional evidence that downstream systems and manufacturing capacity are progressing toward deployment, not the discovery of previously unknown customer relationships.

Near-term trading relevance centers on Broadcom’s next guidance update and evidence that system production schedules support semiconductor shipment assumptions. The larger fundamental benefit falls in 2027 and 2028 as Ethernet scale-up and successive custom accelerator generations expand. The principal risk is correlated dependence on the same deployment schedules: accelerator silicon and networking content are separate revenue opportunities, but they are not independent demand exposures. They can both be delayed by customer power, financing, or system-qualification constraints.


NVIDIA FACES A MORE CREDIBLE CHALLENGE TO INCREMENTAL INFERENCE SPENDING AND FULL-SYSTEM ECONOMICS (READ-THROUGH 2)

Affected company: NVIDIA (NVDA: United States). Direction: negative for long-duration share and pricing assumptions, not evidence of an immediate revenue contraction. Magnitude: strategically large, but the near-term earnings impact is unquantified.

Celestica described an OpenAI system combining custom accelerators, Ethernet scale-up, software, liquid cooling, and rack-scale manufacturing. The network can support a scale-up domain of up to 2,048 accelerators. Management also identified Google TPU programs as a major growth driver and characterized Jalapeño as its most comprehensive and complex design engagement.

The important competitive development is not simply another accelerator chip. It is the emergence of a more complete alternative system capable of addressing the integration challenges that historically reinforced NVIDIA’s position. Commercially useful alternatives require networking, thermal management, software, validation, and reliable volume manufacturing in addition to silicon performance. Celestica’s expanded role provides evidence that these complementary capabilities are becoming available outside NVIDIA’s integrated architecture.

The transmission mechanism is reduced dependence on NVIDIA for selected incremental workloads. Successful alternatives can improve customer bargaining power over accelerator pricing and the associated networking and system package. The potential pressure is greatest where workloads are sufficiently predictable and deployment volumes sufficiently large to justify custom designs. It does not imply equivalent competitive pressure across training, general-purpose acceleration, and every inference workload.

The benchmark evidence must remain narrowly interpreted. OpenAI’s published Jalapeño results showed favorable performance across specified models against selected comparison systems, using accelerator package-power normalization. They did not establish superiority across all workloads or against every subsequent NVIDIA generation. NVIDIA’s latest reported data-center revenue was still growing exceptionally rapidly, and its product roadmap continues to advance.

Near-term catalysts are additional alternative-platform qualification and deployment disclosures, which can affect expectations for NVIDIA’s terminal share before materially affecting current earnings. The substantive earnings test is the 2027 and 2028 allocation of incremental inference capacity. The high-conviction negative conclusion concerns the credibility of competing systems, not a near-term NVIDIA earnings miss or a blanket short thesis.


AMD’S HELIOS EXECUTION RISK IS DECLINING AS THE NETWORKING SUPPLY CHAIN BECOMES MORE CONCRETE (READ-THROUGH 3)

Affected company: Advanced Micro Devices (AMD: United States). Direction: positive. Magnitude: material to the credibility of the data-center accelerator growth trajectory, with limited evidence for changing the immediately upcoming quarter’s estimates.

Celestica’s DS6010 was co-engineered for AMD’s Helios architecture and supports connectivity among 72 MI455X GPUs within a rack. The switch combines 204.8 Tbps of aggregate switching capacity, a high-density cabled backplane, liquid cooling, and SONiC integration. Celestica stated: “We expect to begin ramping by the end of the year.”

This is valuable because competitive accelerator silicon is insufficient if complete systems cannot be qualified, produced, and deployed reliably. Celestica’s disclosure provides evidence of progress in a critical complementary subsystem and its manufacturing process. Successful switch production reduces a potential bottleneck between AMD’s product roadmap and customer deployments.

The longer-duration benefit is an increasingly capable open ecosystem around AMD’s accelerators. Customers gain a more credible alternative to buying an entirely proprietary system, while AMD can rely on specialist partners for selected networking and rack-integration functions. The same disclosure also makes clear that AMD does not capture all system value: Broadcom supplies the switch silicon, and Celestica supplies substantial engineering and manufacturing content.

AMD’s November 3 results are a near-term catalyst for confirming Helios production timing, customer acceptance, and the relationship between initial shipments and meaningful revenue. Celestica’s anticipated year-end switch ramp should not be translated automatically into a large December-quarter AMD revenue contribution. The greater fundamental effect is in 2027.

The key invalidation risk is that networking readiness proves insufficient to overcome constraints elsewhere, including software, accelerator qualification, memory, cooling, or customer-site readiness. The call supports reduced integration risk, not a conclusion that every element of the deployment path is complete.

OPTICAL COMMUNICATIONS

THE 1.6T UPGRADE EXTENDS THE PLUGGABLE OPTICS OPPORTUNITY RATHER THAN IMMEDIATELY DISPLACING IT (READ-THROUGH 4)

Affected company: Coherent (COHR: United States). Direction: positive. Magnitude: potentially large because data-center and communications products constitute the majority of company revenue. Confidence is high in the technology-cycle implication, but Celestica did not identify Coherent as the supplier to a specific awarded program.

The most actionable optical commentary was that “pluggables remain in the 1.6 timeframe,” with CPO expected to become more prominent at 3.2T and 400G SerDes. Celestica simultaneously expects continued 800G growth and 1.6T revenue to overtake 800G in 2027. Its named hyperscaler CPO program is scheduled to ramp only in late 2027.

The implication is a longer and potentially larger commercial window for high-speed pluggable transceivers than a rapid-displacement narrative would suggest. Coherent has identifiable exposure through its 1.6T optical-module portfolio, including 8x200G products. Continued shipments across multiple speed generations can support production scale, manufacturing utilization, and investment recovery before integrated optics becomes more prevalent.

The exposure is financially meaningful. Coherent’s latest reported quarterly data-center and communications revenue was approximately $1.615 billion out of approximately $2.05 billion of total revenue. Celestica’s commentary therefore corroborates an important demand driver rather than merely validating a small development program.

The November 4 earnings report is the next trading catalyst, particularly commentary on 1.6T shipments, manufacturing capacity, yields, and the continuation of 800G demand. This is not a new program award for Coherent and does not establish a specific revenue beat. Optical-port counts, copper connections, supplier allocations, and realized prices are not disclosed sufficiently to construct that bridge.

The longer-duration implication is coexistence rather than an abrupt technology substitution. CPO changes where optical functions reside; it does not eliminate the need for photonics. The risk is that supplier-specific qualification, yield, or component constraints prevent Coherent from converting industry demand into shipments. A favorable market transition cannot compensate automatically for poor execution.

NETWORKING SYSTEMS AND SOFTWARE

CUSTOM SCALE-UP DESIGNS CHALLENGE AUTOMATIC OEM SHARE ASSUMPTIONS, BUT DO NOT ESTABLISH ARISTA DISPLACEMENT (READ-THROUGH 5)

Affected company: Arista Networks (ANET: United States). Direction: negative for assumptions of automatic share capture and proprietary software attachment in the fastest-growing network layers. Magnitude: moderate strategic risk to the AI growth opportunity; no high-conviction near-term earnings reduction follows from this call.

Celestica described its joint design and manufacturing model as an increasingly “preferred approach to product development” for hyperscalers and frontier AI labs. It cited 58% share of custom Ethernet switch revenue and emphasized customized SONiC software across its networking platforms. The Helios and Jalapeño programs demonstrate that sophisticated scale-up systems can be delivered through direct customer, semiconductor, and manufacturing partnerships.

The transmission mechanism is customer procurement shifting toward designs that do not require a conventional branded network-system sale with a proprietary software stack. Ethernet adoption can therefore increase semiconductor and manufacturing revenue without producing equivalent revenue for branded networking vendors. Market growth and vendor share must be modeled separately.

The important counterevidence is Arista’s own October 7 announcement of rack-scale scale-up and scale-out architectures. Its SU-144 offering, cross-rack designs, liquid-cooling capabilities, and support for either EOS or an open network operating system demonstrate an active competitive response. Arista is not confined to traditional front-panel switching or excluded from open architectures.

Accordingly, the high-conviction negative is a limitation on aggressive market-share and software-attachment assumptions, not proof that Arista is losing existing accounts. The upcoming earnings and product-ramp updates are the near-term catalysts for determining which architecture announcements are converting into commercial deployments.

Over 2027 through 2029, the key issue is whether Arista earns differentiated economics from software, operational reliability, and system integration when customers have credible custom alternatives. The competitive concern would weaken if Arista demonstrates substantial scale-up wins with attractive economics. Celestica’s success alone is insufficient evidence for an Arista short position.

POWER DISTRIBUTION AND THERMAL MANAGEMENT

RISING RACK DENSITY CREATES ARCHITECTURE-AGNOSTIC DEMAND FOR ADVANCED POWER AND LIQUID COOLING (READ-THROUGH 6)

Affected companies: Vertiv (VRT: United States) and Eaton (ETN: Ireland). Direction: positive. Magnitude: high fundamental relevance for Vertiv and material relevance for Eaton’s electrical businesses, with a smaller proportional effect on Eaton’s diversified consolidated portfolio.

Celestica is engineering systems approaching 500 kW or more per rack, supporting a transition toward 800 VDC distribution, and evaluating liquid cooling beyond accelerator and switch chips to include memory and optics. Management also described adding hundreds of megawatts of power availability across its manufacturing network.

The important implication is that infrastructure complexity rises across competing compute architectures. These requirements are not dependent on NVIDIA retaining every accelerator sale. AMD, Google TPU, and OpenAI custom systems also require power conversion, protection, distribution, cooling, commissioning, and service. This makes advanced electrical and thermal infrastructure a less architecture-specific exposure than individual accelerator suppliers.

Vertiv’s published 800 VDC work emphasizes new requirements for DC protection, fault management, grounding, maintenance, and service capability. Eaton’s architecture materials similarly address higher-voltage distribution, modular power systems, and cooling integration. These are identifiable capabilities aligned with the requirements Celestica described. Neither company was named as the supplier to a particular Celestica program.

Vertiv’s October 28 earnings report is an immediate trading catalyst for orders, backlog conversion, manufacturing capacity, and cooling-related margins. The longer-duration opportunity lies in 2027 and 2028 deployments and the associated service requirements. Celestica’s production plans support the demand direction but do not establish a specific order award or quarterly earnings surprise.

The economic model should not assume that every component gains content per megawatt. Higher-voltage architectures can reduce current, conductor requirements, and selected conversion stages. Value is likely to migrate toward qualified system solutions, protection, controls, cooling, and service rather than rise uniformly across all legacy products.

Geographically, Celestica’s investment in Texas and Thailand reinforces demand for infrastructure support across both North American and Asian production networks. Factory testing power and end-customer data-center power remain separate requirements. Conflating them would overstate downstream deployment capacity.

CLOUD PLATFORMS AND AI CAPITAL SPENDING

GOOGLE’S TPU ROADMAP GAINS SUPPLY-SIDE VALIDATION, WHILE THE CASH-REINVESTMENT BURDEN REMAINS SUBSTANTIAL (READ-THROUGH 7)

Affected company: Alphabet (GOOGL: United States). Direction: positive for long-term compute availability and strategic control; negative for assumptions of rapid capital-spending normalization and near-term free-cash-flow recovery. Magnitude: material to Google Cloud and consolidated capital allocation.

Celestica identified its “multi-generational relationship with Google in support of their TPU platform” as a major driver of expected custom compute growth. It also stated that awarded-program demand for 2028 is meaningfully above the 2027 revenue outlook. This provides supply-chain corroboration that custom compute investment extends beyond the current product cycle.

The positive transmission mechanism is greater access to workload-optimized infrastructure under Google’s control. Successful TPU production can increase available capacity for internal AI services and external cloud customers while reducing dependence on the commercial terms and deployment schedules of merchant accelerators. More control over hardware, software, and workload optimization can improve long-term unit economics.

The financial timing is less favorable. Hardware delivery generally requires capital spending before the associated cloud revenue and utilization fully mature. Depreciation begins as assets enter service, while power, networking, and operating expenses also increase. Lower cost per unit of computation does not necessarily reduce aggregate spending when usage expands faster than efficiency improves.

Alphabet’s already-reported Q2 results illustrate this distinction. Google Cloud revenue increased 82% to approximately $24.8 billion, with approximately $8.8 billion of operating income. Nevertheless, consolidated capital spending of approximately $44.9 billion exceeded operating cash flow of approximately $39.1 billion, producing approximately $5.9 billion of negative quarterly free cash flow. Celestica’s call corroborates an extended investment cycle; it does not newly reveal those previously reported cash pressures.

The upcoming Alphabet earnings update is the near-term catalyst for capacity availability, capital commitments, utilization, and cash conversion. The longer-duration benefit depends on whether deployed TPU capacity earns adequate returns rather than merely increasing supply. No conclusion follows from this call that Google has raised its formal capital budget by a specific amount.

SEMICONDUCTOR CAPITAL EQUIPMENT

THE ATS ACCELERATION CORROBORATES EQUIPMENT-DEMAND DURABILITY, BUT IS NOT A NEW 40% GROWTH FORECAST FOR EQUIPMENT SUPPLIERS (READ-THROUGH 8)

Affected company: Lam Research (LRCX: United States). Direction: positive. Magnitude: meaningful support for the durability of the equipment cycle, but modest incremental information for near-term estimates because Lam has already reported strong results.

Celestica’s ATS revenue increased 23%, exceeding its mid-teens expectation, primarily because capital equipment accelerated. Management cited forecasts for wafer-fab-equipment spending to grow by more than 40% in 2027, with continued strength in 2028, and described strong relationships with leading equipment OEMs.

The mechanism is straightforward but sequential. More accelerator and memory capacity requires semiconductor manufacturing investment. Greater process complexity also increases the importance of deposition, etch, and other manufacturing steps. Equipment makers then require more components, assemblies, and manufacturing capacity from their supply chains. Celestica’s capital-equipment activity provides corroboration that the investment cycle is reaching manufacturing orders rather than remaining only a demand forecast.

Lam’s October 21 report had already disclosed record quarterly revenue of approximately $8.22 billion and a 40.2% non-GAAP operating margin. Celestica’s call therefore cannot be treated as a prediction of that completed earnings beat. Its incremental relevance is confidence in the persistence of demand into 2027 and 2028.

The quoted industry-growth forecast should not be transferred directly to Lam’s revenue model. Equipment-category exposure, customer geography, service revenue, export controls, manufacturing constraints, and recognition timing determine company performance. Celestica did not identify the OEMs responsible for its specific acceleration, so this is not evidence of a new Lam contract.

The near-term catalyst is subsequent equipment-order and shipment guidance rather than the already-reported September quarter. The longer-duration positive is sustained investment in advanced semiconductor capacity. The principal risk is timing: supplier activity can support equipment already ordered, while cleanroom availability, installation, and customer commissioning still constrain later revenue conversion.

ELECTRONICS MANUFACTURING AND CAPITAL ALLOCATION

AI MANUFACTURING GROWTH SHOULD BE VALUED BY CASH CONVERSION, NOT BY APPLYING A COMMON MULTIPLE TO REVENUE ACCELERATION (READ-THROUGH 9)

Affected companies: Jabil (JBL: United States) and Celestica (CLS: Canada). Direction: positive relative support for Jabil’s cash-generation profile; negative for treating Celestica’s earnings expansion as immediately distributable cash. Magnitude: material to valuation, financing risk, and portfolio construction.

Celestica expects approximately $19 billion of incremental revenue in 2027, $1.5 billion to $2 billion of capital expenditure, and additional working-capital investment. Management stated: “The combination of these investments will result in a net use of cash in 2027.” The target to return to positive cash flow in 2028 remains conditional.

The accompanying filing strengthens the distinction between earnings and internally generated funding. Celestica had $700.8 million of sold receivables outstanding at September 30, while reported Q3 free cash flow was $127.5 million. Supplier liabilities also provided substantial funding. These are legitimate financing mechanisms, but they show that rapid earnings growth does not eliminate reliance on external capital and favorable working-capital terms.

Jabil’s already-published fiscal 2027 outlook presents a different profile: approximately $44.5 billion of revenue, 24% growth, and approximately $1.6 billion of adjusted free cash flow. Jabil is therefore also participating in substantial infrastructure growth while forecasting continued cash generation. Celestica’s cash-use outlook should not be generalized into a conclusion that every AI manufacturing supplier must become cash-flow negative.

The comparison is not perfectly like-for-like. Fiscal periods, product mix, customer terms, capital ownership, growth rates, and free-cash-flow definitions differ. Jabil’s lower growth rate also imposes a smaller incremental funding requirement. The valid implication is that capital efficiency and contract structure deserve a separate valuation assessment from revenue growth and adjusted operating margin.

Near-term catalysts are working-capital movements, receivables-sale balances, customer deposits, and revised capital-spending forecasts at both companies. The longer-duration test is cash retained after growth investment and employee-equity obligations. Celestica could ultimately justify superior economics if its design-led investments generate durable returns, but current ROIC and EPS growth do not establish that outcome in advance.

The geographic read-through is regionalization rather than a wholesale exit from Asia. Celestica is expanding both Texas and its established Asian network. Manufacturing advantage increasingly depends on the ability to combine engineering, available power, regional capacity, and customer qualification. The call does not establish that an unrelated manufacturer will win business simply because it owns facilities in a favored country.

CROSS-PORTFOLIO IMPLICATIONS

The clearest positive read-throughs are Broadcom’s participation across multiple compute ecosystems, Coherent’s extended high-speed pluggable opportunity, and the architecture-independent requirements supporting advanced power and cooling. AMD benefits from more concrete system-readiness evidence. The strongest negative implications concern NVIDIA’s long-term share of incremental inference spending, unearned assumptions of automatic branded-networking participation, and the financing burden of rapid manufacturing expansion.

These exposures remain highly correlated. Broadcom, Coherent, Vertiv, Eaton, Lam, Jabil, and Celestica can all benefit from the same infrastructure cycle, but they can also be affected simultaneously by deployment delays or weaker capital availability. Owning different layers of the supply chain does not necessarily diversify the underlying spending factor. Likewise, stronger supplier revenue can coincide with weaker customer free cash flow before the installed infrastructure produces adequate returns.

No company-specific upgrade is warranted from this call alone for unnamed memory, substrate, connector, automation, defense, or healthcare suppliers. The transcript does not establish the supplier allocations or revenue exposure required for high-conviction conclusions. The retained findings are grounded in the October 27 Celestica call, its accompanying financial disclosures, and the relevant companies’ published earnings and product materials available by that date.