Neutral
AMATApplied Materials, Inc.Applied Materials, Inc.
MSFTMicrosoftMicrosoft
NVDANVIDIANVIDIA
PANWPalo Alto NetworksPalo Alto Networks Weekly|The Rebound and What Comes Next, NVDA Financing, Cybersecurity, CPO, IFX & AMAT
Last week, we argued that the medium-term questions facing AI could coexist with a near-term rebound after the sharp selloff. In the first half of the week we saw a sharp rebound, which was later interrupted. We would keep the same distinction between time horizons: a recovery in share prices does little to resolve the gap between ambitious infrastructure spending and the revenue needed to support it. Jensen Huang’s departure adds another reason to look more carefully at how that spending gets financed. As we discuss in this week’s NVDA deep dive, Huang’s role extended well beyond selling GPUs. NVIDIA’s equity investments, capacity backstops and credit support helped Neolabs and Neoclouds raise capital and commit to expansion. A more conservative approach under new leadership could make the medium-term constraints we discussed last week more visible. The first changes may show up in new investments and financing commitments over the next one or two quarters, with consequences for deployment and hardware demand emerging later. Strong near-term GPU demand can coexist with a less supportive funding environment. We think the market needs to scrutinize who is financing the next round of capacity, on what terms, and how much spending depends on continued support from NVIDIA. This week also brought more evidence of the security and operational risks surrounding increasingly connected software. Fireworks’ security incident exposed data from some Notion workspaces, illustrating how an AI infrastructure vendor’s breach can reach customers of the applications built on it. Separately, the prolonged Microsoft 365 outage disrupted essential business services. Microsoft attributed the outage to a network upgrade issue exposed by a subsequent change, and said it was unrelated to cyberattacks or AI/agents. It does, however, underline how much business activity depends on a small number of shared services. These developments reinforce our focus on the cost of securing and maintaining the systems that AI increasingly relies on. As AI expands software development and accelerates vulnerability discovery, enterprises face more work in remediation, identity management, third-party oversight and continuous protection. We see a firmer basis for sustained cybersecurity demand, even while the funding and returns on parts of AI infrastructure remain uncertain. The timing still matters: our latest checks point to security-spending growth among covered customers rising from 10–12% in 2026 to 13–15% in 2027, with much of the immediate response funded through existing budgets. Our PANW preview provides an early read on how that demand is reaching established security platforms. This Week’s Reports NVDA: Near-term demand meets a more uncertain financing backdrop. Our earnings preview examines the B300 ramp and Rubin shipment timing, while our new deep dive considers how Huang’s departure could change NVIDIA’s support for Neolabs, Neoclouds and its supply chain. CPO: Bonding and testing still limit the ramp. Our equipment and process checks point to large-scale shipments in late 2027–2028, with hybrid-bonding yields and an unresolved test-vendor decision keeping near-term volume expectations in check. Cybersecurity & PANW: The budget response is taking shape. Our latest cybersecurity checks point to faster spending growth in 2027, while the PANW preview examines solid customer spending and continued consolidation onto fewer security platforms. Premium Report Snapshot A portion of our research is reserved for Premium subscribers and is not distributed via Substack. Below is a snapshot of what Premium subscribers received this week beyond the Substack feed. Subscribe now Read more
View original →Neutral
AMATApplied Materials, Inc.Applied Materials, Inc.
FORMFormFactorFormFactor
NVDANVIDIANVIDIA
STXStacksStacks
TELTE ConnectivityTE Connectivity
TERTeradyneTeradyne
TSMTSMCTSMC
XPLPlasmaPlasma Deep | What’s Holding Back Co-Packaged Optics Mass Production?
Why this note matters This note is for investors following the co-packaged optics (CPO) ramp and the companies exposed to it: hybrid-bonding tool makers (Besi, Tokyo Electron, EV Group), CPO test equipment vendors (ficonTEC, MPI, Teradyne, Advantest, FormFactor, Chroma), and the foundry and OSAT supply chain (TSMC, ASE/SPIL). We interviewed several equipment and process experts on where CPO mass production stands, with a focus on hybrid bonding, the step that currently limits yield. The pace of scale-out CPO also matters for NPO: the longer CPO takes to mature, the more likely NPO is to be used in scale-up, and for longer. We then walk through the open issues in each of the four test insertions, our checks on the main vendors at each step, and a bottom-up estimate of the CPO test equipment TAM for 2027-2028. This Substack is reader-supported. To receive new posts and support my work, consider becoming a free or paid subscriber. Conclusion 1: Compared with our June note on the CPO delay debate and our October OCP summary, we are now more cautious on the pace of scale-out CPO. CPO for scale-out switches is in production and shipping in small volumes, but the ramp is limited by bonding and testing, and large-scale delivery will come in late 2027 to 2028. CPO volume production still faces two main hurdles: 1) the W2W tools for hybrid bonding are not ready; and 2) an equipment vendor has not yet been selected for Insertion 2 testing. Bonding yields on the TSMC COUPE line, the open Insertion 2 vendor decision and customers adjusting their timelines mean 2027 deliveries will fall short of the roughly 100k Spectrum-X CPO switch target cited in our OCP note, with volume shipments pushed to late 2027 to 2028 (later than the market expects). This report covers the optical engines used in NVIDIA’s COUPE-based scale-out CPO switches. Our view on scale-up is unchanged: NVIDIA’s next-generation scale-up will certainly use optics, but whether that is CPO or NPO is not yet decided, and the constraints described here make an early move to CPO in scale-up less likely. Demand for optical interconnect and high-power external light sources is not affected; what changes is the timing. (I) Direct cause of the CPO delay #1: hybrid bonding This report uses the 3D hybrid-bonded optical engine in Spectrum-X CPO as its main case. Wafer-level hybrid bonding on the COUPE platform works much like SoIC. It has two steps, D2W (current yield 80%) and W2W (current yield 50%), and neither yet meets the yield needed for stable, profitable mass production (each step needs to exceed 90%, and W2W is the step that most needs to improve). The main bottleneck today is that W2W tools with a vacuum chamber, which could lift W2W yield to about 90%, have not been successfully developed yet. These tools are expected to sample in 2H27, with the vendor decision in 1H28. This is one of the direct causes of the CPO delay. Figure 1: TSMC CoWoS platform with SoIC logic stacks and COUPE optical engines Source: TSMC public materials, redrawn by FUNDA How hybrid bonding works: Hybrid bonding forms the dielectric bond and the Cu-Cu metal bond at the same interface in one step. It is a bumpless direct-bonding technology. Compared with conventional micro-bumps at a 40-50 µm pitch, hybrid bonding can shrink the interconnect pitch below 10 µm, giving much higher I/O density over a shorter electrical path. The standard flow is: (1) CMP the bonding surfaces so the dielectric (SiO2/SiCN) is extremely flat, with the Cu pads recessed slightly below it by a few nanometers; (2) clean and plasma-activate the surfaces to raise bond energy; (3) align and bring the surfaces into contact at room temperature, so the dielectric bonds first; (4) anneal at about 200-400°C, where the Cu expands to fill the recess and forms a metallic Cu-Cu bond; (5) after bonding, use infrared or acoustic scanning (C-SAM) to look for voids and unbonded areas. In TSMC’s SoIC and COUPE flows, D2W and W2W are two consecutive steps in the same flow, not two competing technology routes. How the two steps fit together: In the COUPE flow, D2W first places screened, known-good EIC dies and lenses one by one onto the PIC wafer. W2W then bonds the whole wafer to a carrier wafer, which supports the later flip, thinning and backside routing. The expert we spoke with has hands-on bonding process experience and sees the two steps as consecutive stages of one flow, with W2W partly serving as temporary support for the backside process. Expert A (bonding process expert at an equipment maker): “Die-to-wafer and wafer-to-wafer are often used together in the same product flow. For example, if you need backside routing later, you first do a temporary wafer-to-wafer bond to hold the structure.” The yields of the two steps therefore multiply, and a weak step drags down the final output of the whole wafer. That is why we use compounded yields when sizing the equipment for both steps below. (1) Step 1: Die-to-wafer (D2W) This step covers EIC-to-PIC hybrid bonding and OE-to-lens bonding. The technology is relatively mature. The tool vendor is Besi, at roughly US$3mn per tool, and Besi is close to a sole supplier. Assuming a 90 mm² OE die, a 12-inch wafer yields about 700 dies. Bonding takes 2-3 hours per wafer, so one D2W bonder at 100% utilization can process about 2.04mn OEs a year (700 dies per wafer × 8 wafers per day × 365 days). Current yield at this step is 80%. Process flow: The EIC wafer is tested and diced first, and only known-good dies (KGD) are kept. After the dies and the PIC wafer go through CMP and plasma activation, the bonder picks each die, flips it, aligns it to its target site on the PIC wafer and places it, forming a pre-bond at room temperature. Once the wafer is fully populated, it is annealed in one batch to form the Cu-Cu bonds. Lens dies are placed the same way, one at a time. What this step does well and where it falls short: D2W lets you bond only good dies and allows dies of different sizes to be integrated together, which helps overall yield and cost. That is why steps like EIC and lens attach, where you want to screen before bonding, use D2W. Placing dies one at a time limits throughput. Particles and die-edge defects from dicing also require extra cleaning. Besi is the main supplier of D2W hybrid bonders and develops integrated solutions together with Applied Materials. Capacity and tool supply: The core value of D2W is screening first and bonding second. The expert sees the ability to pick good dies before bonding as the main yield and cost advantage of D2W. Expert A (bonding process expert at an equipment maker): “The biggest advantage of die-to-wafer is that you can pick out the good dies first and then bond them. That makes a very clear difference to overall yield and cost efficiency.” On capacity, the expert estimates that D2W bonding of a 12-inch wafer takes about 2-3 hours, so one tool handles about 8 wafers a day. The foundry plans for roughly 100 D2W tools and has about 30 today. Roughly 30 tools translate into about 7,300 wafers a month. The expert also flagged Besi’s delivery capacity as a possible constraint on expansion. Expert A (bonding process expert at an equipment maker): “The foundry needs a lot of Besi tools. I think Besi’s delivery capacity could become a bottleneck.” Figure 2: COUPE die-to-wafer bonding: placing EIC and lens dies on the PIC wafer Source: Expert interviews, foundry public materials, industry process-flow materials, redrawn by FUNDA Based on Table 1, shipping 100k switches in 2027 would require about 14 Besi tools on site at the foundry, assuming 50% utilization (CPO is still early and not yet in volume production, and bonding tools are ramping, so utilization is low). Our channel checks suggest the foundry already has enough tools on site. If 2028 demand reaches 450k switches and W2W yield (50%) and back-end test and assembly yields stay unchanged, about 61 D2W tools would be needed, and Besi’s delivery pace becomes a factor to monitor. If yields after D2W improve in 2028, theoretical CPO delivery guidance would rise as well, so Besi’s delivery capacity remains a constraint worth tracking. Table 1: D2W equipment requirement Source: Channel checks, FUNDA internal estimates (2) Step 2: Wafer-to-wafer (W2W), the main hybrid-bonding bottleneck This step bonds the carrier wafer to the lens surface. Because the lens surface is not flat, gas tends to get trapped between the lens and the carrier. Tool vendors need to supply bonders with a vacuum chamber to prevent voids, but those tools are still in development. This is where the biggest yield problem in hybrid bonding sits. Figure 3: COUPE wafer-to-wafer carrier bonding and the risk of voids Source: Expert interviews, industry process-flow materials, redrawn by FUNDA W2W still runs on non-vacuum tools today, with a low yield of about 50%. EV Group (EVG) and Tokyo Electron (TEL) are the vendors able to design vacuum-chamber tools, and TEL is seen as being further ahead. We expect early demo tools in 2H27 and mature tools in 2028 . A tool is expected to cost about US$30mn (an expert estimate, since vacuum-chamber tools have not yet been successfully developed; the price should come down once in volume production). Throughput of the vacuum tool cannot be estimated yet. Taking the non-vacuum tool as a reference at about 20 wafers per hour, one tool can process about 122.6mn OEs a year (20 wafers/hour × 24 hours × 365 days × 700 dies). At 50% utilization, with 50% yield at this step and 65% downstream yield, one tool can support about 550k CPO switches. Process flow: W2W aligns two full wafers and bonds them in one shot. In the COUPE flow, after D2W attaches the EICs and lenses, the whole wafer is bonded to a carrier wafer. The carrier gives mechanical support for flipping, backside thinning, TSV reveal and bumping, before dicing and 2.5D packaging. Our channel checks indicate the new platform has settled on silicon carbide (SiC) carriers. What hurts yields: The lens surface is not flat. When it is bonded to a flat carrier, gaps form at the interface and trap gas, leaving bubbles or unbonded patches. Bubbles can also expand during anneal and spread from one die to its neighbors. A vacuum chamber pumps the gas out before contact and is the main way to cut voids. Carrier bonding bonds the whole wafer at once, so throughput is high, but one void can affect many dies on the wafer, and the step is more sensitive to warpage and particles. Yield at this step directly sets the output of the whole wafer. The expert believes the EIC-PIC bonding process itself is largely proven. The main mass-production challenge is bonding a non-flat surface like the lens to a flat carrier. Expert A (bonding process expert at an equipment maker): “EIC-to-PIC bonding and the related processes have basically been verified. Different ICs and substrates may need recipe tweaks, but that is not the main problem now. The harder part is bonding an uneven structure like the lens to a flat carrier.” The expert then explained how voids grow from a single-die issue into a wafer-level yield problem, and gave a rough estimate of yield below 50%. This is a judgment, not measured data. Expert A (bonding process expert at an equipment maker): “Whether the lens is concave or convex, there will be gaps at the interface, and you get bubbles or areas that do not bond. If the bubbles or particles expand during anneal, a defect can spread from one die to several, and yield can fall below 50%. For volume production you normally need at least around 90%.” Timeline for vacuum bonders: Bonding in a vacuum reduces trapped gas at the interface, but there is no proven production tool yet. The expert gave several ranges for the development timeline, all with wide uncertainty. Expert A (bonding process expert at an equipment maker): “For the new generation of vacuum tools, I think it will take at least two to three years, or a year and a half to two years if things go fast. We might see some demo results by the end of 2027, and it could happen in 2028, but none of this is certain.” On vendors, our channel checks show that EVG already has a relevant platform, but the foundry has concerns about its bonding alignment and leans toward evaluating TEL, with EVG as a backup. The foundry also wants a second source, though other vendors do not yet seem to match that performance. If vacuum tools are not ready in time, the process sequence may need to change, for example by modifying the wafer bumping steps or doing some steps on the carrier earlier. Whether these alternatives work, and which one the foundry will choose, is still unclear. Inspection: Before bonding, high-resolution optical scans can catch particles of 1-2 µm. After bonding, acoustic scanning (C-SAM) checks for voids and unbonded areas. Inspection only finds defects; it cannot replace control of where the defects come from. Expert A (bonding process expert at an equipment maker): “Inspection can find the defects, but it cannot replace controlling the sources of those defects, like particles and bubbles.” So W2W yield improvement depends mainly on the tools and the process. Adding inspection only keeps more bad parts from reaching the back end; it does not raise bonding yield. W2W equipment requirement (utilization also assumed at 50% during ramp): Table 2: W2W equipment requirement Source: Channel checks, FUNDA internal estimates Post-bond steps: After W2W, the wafer is flipped and the carrier supports the stack. The device side is back-ground and thinned by CMP; once the TSVs/TDVs are revealed, the redistribution layer and bumps are built. The wafer then goes to the OSAT for dicing, attachment to the package substrate and FAU coupling, followed by Insertion 3 and Insertion 4 testing. Figure 4: COUPE post-bond processing: flipping, thinning and bumping Source: Expert interviews, industry process-flow materials, redrawn by FUNDA (3) Summary: how bonding sets the pace of CPO mass production Overall, D2W tools and the process are relatively mature, and near-term capacity can cover 2027 demand. Low bonding yields raise product cost, but they have not stopped shipments. What really sets the pace of volume production is W2W carrier bonding: non-vacuum tools yield about 50%, vacuum tools should only reach early samples in 2H27, and they may not mature until 2028. This matches what we hear from the test side. The test expert we interviewed believes low yields and integration difficulty have already pushed some customers to adjust their adoption timelines. Expert B (test equipment expert): “Wafer-to-die yield is only about 30% to 40% right now (the combined yield from the PIC wafer to the final good optical engine, not the yield of the D2W step alone), and integration is hard, so enthusiasm for CPO has cooled recently. Several customers are adjusting their timelines and pushing full CPO adoption to 2028.” Subscribe now Read more
View original →