Marvell Technology guided FY28 revenue to roughly $20 billion. Three years ago, the company printed about $5.5 billion. That is not a growth rate. That is a phase transition, and phase transitions do not happen because a spreadsheet says they should. They happen because something underneath the system changes state. So before I accept the number, I want to trace the gas trail back to the genesis block, because the headline tells you what management hopes the demand curve does, and says almost nothing about whether the silicon, the packaging, the foundry allocation, or the customers will allow it to happen.
The framing of the original report was a title-level flash: AI and connectivity demand rising, FY28 outlook raised to ~$20 billion. Five information points, everything else left for the analyst to reconstruct. That is the condition I work in. When a press release hands you a terminal value and hides the state machine, your job is to rebuild the state machine. What follows is that reconstruction โ a line-by-line audit of a fabless designer that has quietly become the second-most-important custom-silicon house on earth, and the supply-chain, geopolitical, competitive, and accounting assumptions buried inside a single forward-looking number.
I should be explicit about confidence levels up front, the way I would in a formal vulnerability report. Design capability: 6/10. Supply chain positioning: 6/10. Capacity and capex: 5/10 โ the weakest, because fabless companies do not disclose foundry allocations and the market consistently mistakes demand for capacity. Demand structure: 6/10. Geopolitics: 6/10. Competitive dynamics: 6/10. Financials: 5/10. Nothing here is a certainty. Everything here is an invariant test. Entropy increases, but the invariant holds โ and the invariant I am testing is whether a demand prediction has been mislabeled as a capacity commitment.
Context: What Marvell Actually Sells
To audit the number you first have to know what the company is. Marvell is fabless. It owns no fabs. It does not purchase lithography tools, it does not run a yield learning curve on its own process, and it does not carry the depreciation burden of a foundry. That single structural fact reframes every dimension that follows. When someone asks about Marvell's "process technology," the correct answer is that Marvell's process technology is a contract. It is Taiwan Semiconductor Manufacturing Company's N5, N4, N3, and eventually N2 nodes, purchased as a service, allocated by negotiation, and constrained by capacity the customer does not control.
What Marvell actually owns is design capability, and it owns it in two distinct engines that the market persistently conflates into one AI narrative. The first engine is custom silicon โ the XPU and ASIC business, where hyperscale cloud providers commission application-specific accelerators rather than buying general-purpose GPUs. The second engine is connectivity, and specifically the optical and high-speed interconnect silicon that moves data between and inside those accelerators. The connectivity engine is the Inphi inheritance. The custom-silicon engine sits on top of an Arm architecture license inherited from the Cavium acquisition, expressed today in the Octeon line of Arm-based server processors and DPUs.
This dual structure matters because the two engines respond to different physical constraints. Custom ASIC revenue scales with the number of accelerator projects a hyperscaler decides to tape out. Optical DSP revenue scales with the size of the cluster โ with the number of links required to wire thousands of accelerators into one coherent training or inference fabric. The larger the AI cluster, the more rigid the demand for optical interconnect becomes, because you cannot build a ten-thousand-node fabric without a proportionate explosion in high-speed SerDes channels and PAM4 optical DSPs. Marvell is the global leader in that optical DSP segment. That is not a marketing claim; it is a market-share fact I will defend in the competition section.
So when the headline attributes the raise to "AI and connectivity," it is really naming two businesses with two different risk profiles and two different moats. The custom ASIC business is high-revenue, high-visibility, and structurally fragile because the customer can internalize it. The connectivity business is lower-profile, deeply technical, and structurally durable because optical DSP requires analog and mixed-signal accumulation that is extremely difficult to replicate in-house. One of these is a bet. The other is an annuity. The $20 billion target leans on both, but the market prices it as if it were all annuity.
The Design Layer: Where the Moat Is and Where It Isn't
Let me start the technical dissection where a forensic auditor starts โ with the primitives. Marvell does not fabricate transistors; it designs chips and licenses intellectual property. The most valuable of that IP is the SerDes block. SerDes โ serializer/deserializer โ is the circuitry that takes parallel data inside a chip and turns it into a high-speed serial stream for transmission across a link. Marvell's 112G and 224G SerDes IP sits at the center of AI cluster interconnect. At 224G, the signal integrity problem becomes brutal: insertion loss, crosstalk, jitter, and equalization all fight each other, and the analog design skill required to win that fight is scarce. Marvell and Broadcom are the two houses that reliably ship it. Everyone else is either a fast follower or a customer.
This is where I apply a lesson from my own audit history. In 2018, as a junior analyst in Madrid, I spent three months inside the 0x Protocol v2 contracts, ignoring the business logic entirely to reverse-engineer the Order Manager's assembly. I found seven edge cases in the signature-verification path that everyone else had walked past, because they were reading the whitepaper and I was reading the opcodes. The lesson was not that assembly is magic. The lesson was that the moat is never in the part of the system the marketing describes; it is in the part of the system that is hard to copy. For Marvell, the marketing describes the AI accelerator. The moat is in the SerDes and the optical DSP โ the parts nobody writes a press release about.
Now the honest accounting of the design layer. On the custom ASIC side, Marvell is the clear global number two behind Broadcom. The gap is roughly half a body โ not a generation of silicon, but a generation of customer scale and design-win count. Broadcom has accumulated more hyperscaler programs, more tape-outs, and more accumulated design-service relationship capital. Marvell's technical capability in a single custom project is competitive; its capacity to run many simultaneous projects at Broadcom's scale is not. This is a distribution problem disguised as a technology problem โ and it is the same pattern I have written about repeatedly in the Layer 2 wars, where the real difference between OP Stack and ZK Stack was never the cryptography but who could convince more teams to deploy chains first. Custom silicon is the same game with different physics. Whoever lands the most design wins compounds the relationship, the IP reuse, and the tooling advantage. Technology parity does not win that race; installed base does.
On the optical DSP side, the story inverts. Marvell is number one, with a share north of fifty percent in PAM4 optical DSP, ahead of Broadcom and Cisco. This is the Inphi legacy, and it is the deepest asset in the portfolio. The reason it is deep is that optical DSP requires a blend of high-speed analog, digital signal processing, and mixed-signal design that cannot be conjured by a hyperscaler's internal chip team in a reasonable timeframe. A cloud provider can build a digital accelerator. Building a world-class optical DSP from scratch is a decade-long endeavor. That asymmetry is the structural reason the connectivity engine is more defensible than the accelerator engine, and it is the reason I would weight the durability of the $20 billion number toward the connectivity half.
There is a hidden implication here that the original report missed entirely. If Marvell's AI revenue is genuinely a two-wheel structure โ custom logic plus optical connectivity โ then its revenue is bound to the scale-out of AI clusters, not merely to GPU unit shipments. Those are different exposures. A world where GPUs ship in flat volumes but clusters keep growing in size still expands optical interconnect demand. A world where clusters consolidate onto fewer, larger fabrics still favors the interconnect leader. Marvell is levered to cluster topology, not just to accelerator count, and that is a structurally favorable place to be levered. The connectivity wheel turns even when the accelerator wheel stalls.
The Supply Chain Is the Ceiling
The most important sentence in this entire analysis is the one the original report did not write: Marvell's $20 billion target is a demand forecast, not a capacity commitment. Because Marvell is fabless, it cannot unilaterally deliver a single wafer of that revenue. Every dollar of custom ASIC revenue depends on two externally controlled resources โ TSMC's advanced-node allocation and TSMC's CoWoS advanced-packaging allocation โ and both are contested.
CoWoS โ chip-on-wafer-on-substrate โ is the packaging technology that lets high-bandwidth memory sit alongside a logic die on a silicon interposer. Without it, an AI accelerator cannot be assembled. And CoWoS capacity has been the single hardest bottleneck in the entire AI supply chain. The demand for CoWoS has run structurally ahead of supply, and the allocation is prioritized toward the largest, most strategically important customers. That means NVIDIA first, Broadcom's hyperscaler programs second, and everyone else in the queue. Marvell is not NVIDIA. Marvell is not Broadcom. When CoWoS is scarce, Marvell competes for the remainder, and the remainder is exactly where delivery schedules slip.
The dependency table is unforgiving. On manufacturing, Marvell depends almost entirely on TSMC's leading nodes, with Samsung as a theoretical alternative that has historically lagged on both yield and capacity. On packaging, it depends on TSMC's CoWoS, with OSAT providers like ASE and Amkor offering advanced packaging that is not yet at the required scale for the most demanding AI parts. On architecture, it holds an Arm license, with RISC-V as an unexercised option it has not pursued at scale. On memory, it depends on HBM supplied by the customer or system integrator โ SK Hynix, Samsung, or Micron. The only link in the chain Marvell fully controls is its own SerDes IP. Everything else is a negotiation.
This is where I want to bring in a model from my own work that maps almost perfectly. In 2024, I spent two weeks modeling the economic security thresholds of the EigenLayer restaking architecture, and I published simulation scripts demonstrating that the slashing conditions for active validators were too loose relative to the economic stake required to attack them. The core insight of that analysis was not that restaking is broken. The insight was that a system's security is set by its weakest external dependency, not its strongest internal design. Marvell's situation is the same shape. Its internal design capability is elite. Its external dependency โ a single foundry and a single packaging line, both concentrated in one geographic region โ sets the actual ceiling on what it can ship.
There is a tail risk here that neither the headline nor the original report touched, and it is the largest single risk in the entire thesis. The manufacturing and packaging are geographically concentrated in Taiwan. A disruption there does not selectively hurt Marvell; it hurts the entire industry simultaneously. But when capacity is being restored and reallocated, the recovery order follows strategic priority. NVIDIA and Broadcom get served first. Marvell, smaller and less systemically critical to the hyperscalers' GPU fleets, waits. The tail risk is not that Marvell loses a customer to a competitor; it is that in a systemic shock, Marvell is structurally last in the queue to recover. That is a risk you cannot hedge with a design win.
The Capacity Illusion
I want to dismantle a category error that the market makes constantly with fabless companies, because it is the analytical root of most overvaluation in this sector. When analysts model a fabless company's growth, they frequently treat revenue as a function of demand. Demand is necessary but not sufficient. Revenue is the minimum of demand and deliverable supply. For a foundry, deliverable supply is a function of capex โ you can literally see the fabs being built. For a fabless designer, deliverable supply is a function of someone else's capex, allocated by someone else's priority list. Marvell's capacity utilization is not a number Marvell reports; it is a number TSMC decides.
The capex structure, though, is where the fabless model earns its keep. Marvell's capital expenditure as a percentage of revenue runs in the single digits โ a fraction of the thirty-to-forty-five percent range that foundries carry. That means minimal depreciation drag, high return on invested capital, and a structurally lighter balance sheet. In a normal semiconductor cycle this is an unambiguous advantage. In an AI cycle it becomes subtler. The light-asset model means Marvell cannot build its way out of a supply constraint. It can only design its way into better allocation priority by being a customer TSMC cannot afford to deprioritize.
The timeline mathematics deserve scrutiny. A custom silicon program moves from design win to volume production in roughly eighteen to thirty-six months. That means the $20 billion FY28 target, landing around January 2028, implies that a cluster of large custom programs must already be in design or early ramp today, in FY26 and FY27, and must all reach volume in a narrow window. There is no scenario in which a program signed in FY27 contributes meaningfully to FY28 revenue. So the target is effectively a statement about programs that are already substantially committed. That cuts both ways. It means the target is less speculative than it looks on the demand side โ the customers have already chosen Marvell for those sockets. But it also means the target has almost no room for slippage. If even one anchor program slips a quarter or two, the number does not bend; it breaks.
And the ramp is not solely Marvell's to control. The final deployment pace is set by the hyperscaler's own data-center buildout schedule, its power and cooling availability, and its internal software readiness to actually run production workloads on the new silicon. Marvell can ship the chip. It cannot force the customer to deploy it on schedule. Full production is a joint venture between a fabless designer and a customer who is simultaneously a potential competitor โ a strange and fragile arrangement I will return to, because it is the deepest blind spot in the entire bullish case.
Demand: The Inference Thesis and Its Two Fault Lines
The demand case for the $20 billion number rests on a specific and, I think, largely correct thesis: the largest opportunity in custom AI silicon is inference, not training. Training is dominated by the largest-scale, most performance-sensitive workloads, where general-purpose GPUs retain a strong hold because of their flexibility and their software ecosystem. Inference is different. Inference is a cost and power problem at enormous volume. For a hyperscaler running inference at scale, an application-specific accelerator with a better total cost of ownership and a better performance-per-watt ratio is worth commissioning, even at significant design cost, because the savings compound across billions of queries.
That logic is sound, and it explains why hyperscalers are willing to fund custom programs at all. It is also why Marvell's custom ASIC business has a real and growing addressable market. But I want to test the thesis against its two fault lines, because the headline treats the demand curve as monotonic and it is not.
Fault line one is customer concentration. The custom AI business is concentrated in a handful of hyperscale buyers, and the top few customers likely represent the overwhelming majority of that revenue, with a single largest customer potentially accounting for a third or more of AI revenue. The original report correctly flagged that "dependence on large technology companies carries significant risk," but it did not diagnose the actual shape of the risk. The risk is not customer churn in the ordinary sense. The real risk is that the customer becomes the competitor. Every hyperscaler large enough to commission a custom accelerator is also large enough to build an internal chip team. And several already have. The internal design organizations at the major cloud providers are not a temporary experiment; they are a permanent structural feature of the industry, and they are the most lethal long-term threat to Marvell's custom business. A customer who internalizes a program does not churn โ they simply stop commissioning the next one, quietly, and the revenue does not come back.
Fault line two is the nature of the demand signal itself. AI-related demand is in a build-and-expand phase, and AI silicon is genuinely supply-constrained. But Marvell's traditional businesses โ carrier infrastructure and enterprise networking โ are in a different phase entirely, still working through inventory normalization. The $20 billion target is carried almost entirely by the AI half; the traditional half is a drag, not a driver. That concentration is fine as long as AI capex keeps compounding. It is catastrophic if the hyperscalers decide, for any reason โ power constraints, capital discipline, model-efficiency gains that reduce compute intensity โ to slow the buildout. And here is the uncomfortable truth about AI capital expenditure: it is discretionary. It can be deferred. Unlike a regulatory mandate, it can be cut in a single budget cycle.
On pricing, the picture is mixed. Custom ASIC pricing power is real, because the customer is buying TCO reduction and will pay a premium for it. But that pricing power is bounded by the customer's credible threat to self-design. Meanwhile, the foundry is raising prices on leading-edge nodes, which compresses fabless margins from the cost side. The net effect is that Marvell's gross margin is pulled up by product mix โ more AI, less legacy โ and pulled down by rising wafer costs. The two forces are roughly in tension, which is why I expect margin to drift up modestly rather than expand dramatically.
Geopolitics: The Net Beneficiary Nobody Models
Export controls are usually framed as a headwind for semiconductor companies, and for many they are. For Marvell, I think the net effect is closer to a tailwind, and the original report missed this entirely. Marvell is not on any entity list. But its advanced AI products fall under US export-control constraints on advanced computing and high-performance interconnect, which means it cannot sell its most advanced AI silicon into China. On its face, that is lost revenue. In practice, Marvell had already been structurally reducing its China AI exposure, and its China revenue share has been declining. The advanced AI products it cannot sell to China are not products it was banking on selling to China.
Meanwhile, the geopolitical logic cuts the other way in a way that is genuinely favorable. The hyperscalers that buy Marvell's custom silicon have two simultaneous strategic imperatives: reduce dependence on a single dominant GPU vendor, and localize their supply chains. Both imperatives point toward cultivating Marvell as a credible second source. A US-designed custom accelerator, manufactured by a US-allied foundry, fits the localization agenda perfectly. Marvell is a net beneficiary of supply-chain nationalism, not a victim of it โ because its customers are the very institutions driving that nationalism, and they need a second vendor to make it real.
The China demand-side picture is largely a wash for Marvell specifically. China's export controls on gallium, germanium, and rare earths raise global supply-chain costs, but the effect on a fabless designer is indirect and limited. China's domestic semiconductor push accelerates import substitution in Marvell's traditional China business โ carrier and enterprise networking โ but that business is no longer the growth engine. The material conclusion is that Marvell's geopolitical exposure is asymmetric: modestly negative on the traditional side, modestly positive on the AI side, netting out somewhere between neutral and positive.
But I want to end this section with the risk the report never named, because it is the one that actually keeps me up. Marvell's largest geopolitical tail risk is not the US-China trade relationship. It is Taiwan. The manufacturing and packaging concentration in a single geography is the systemic vulnerability, and it dwarfs any export-control question. This is the same class of risk I flagged in my L2 scalability work, where the fraud-proof bond sizes in early optimistic rollup iterations were mathematically insufficient to deter a sophisticated attacker. The lesson there was that a system can look robust under normal conditions and be catastrophically fragile under adversarial ones. Marvell looks robust today. Its fragility is geographic and it is unhedged.
Competition: Broadcom Is Not the Real Threat
The competitive map is clear, and the original report drew it correctly, if incompletely. In custom AI ASIC, Broadcom leads with roughly sixty to seventy percent share and Marvell sits second at roughly twenty to twenty-five percent. In optical DSP, Marvell leads with more than half the market, ahead of Broadcom and Cisco. In high-speed SerDes IP, Marvell is in the leading tier alongside Broadcom. In storage controllers, Marvell is number one. So the picture is of a company that is second in the largest market and first in several adjacent ones.
The research-intensity comparison is where scale becomes visible. Broadcom spends multiples of what Marvell spends on research and development in absolute terms. Marvell's research intensity is high โ roughly a quarter to a third of revenue, reflecting heavy custom-program investment โ and its research efficiency is genuinely good because it focuses on high-value niches. But efficiency does not fully offset scale. Broadcom can run more large custom programs in parallel than Marvell can, and in a land-grab market, parallel capacity compounds. This is the same dynamic that decides Layer 2 outcomes: not which chain has the better cryptography, but which ecosystem can absorb more simultaneous deployments. Marvell's technical parity in any single project does not translate into victory across many simultaneous projects when the competitor can field more teams.
But here is the contrarian turn. Broadcom is the visible competitor, and the market watches that race. The invisible competitor is the customer. The five-forces picture is brutal precisely because of this: buyer power is strong because the buyers are a few trillion-dollar firms that can self-design; supplier power is strong because the foundry and the packaging line allocate scarce capacity by priority; the threat of new entrants is moderate from design-service houses but severe from the customers' own internal teams. Marvell is squeezed from both ends. It sits between a supplier it cannot outrank and buyers who can, in principle, replace it. Its bargaining power is structurally weak even as its technical position is strong.
This is why I argue that the deepest moat Marvell owns is not the custom ASIC franchise at all. It is the optical DSP franchise โ the connectivity wheel. That business is number one globally, and critically, it is the piece of the portfolio that is hardest for a hyperscaler to internalize. A cloud provider can build a digital accelerator. Building a leading optical DSP requires deep analog and mixed-signal expertise accumulated over years, and the ROI on building it in-house is poor compared to simply buying the best part. So the connectivity business is defensible in a way the accelerator business is not, and I suspect it is the most undervalued component of Marvell's valuation. The market prices the exciting, fragile business and underweights the boring, durable one.
The Arithmetic of a Fifty Percent CAGR
Now the financial skeleton, and I will be candid about the confidence level here because the original report provided no financial data and I am reconstructing from public information. Take a FY25 revenue base in the range of five and a half to five and eight billion dollars. Set FY28 against a roughly twenty-billion-dollar target. The implied three-year compound annual growth rate is on the order of fifty percent. Fifty percent. In the history of the semiconductor industry, sustained growth rates at that level are essentially unheard of outside of the earliest ramp of a genuinely new product category. A fifty percent CAGR is not a forecast; it is a bet that AI custom silicon is a new category large enough to rewrite the industry's growth math.
Gross margin is currently in the low sixties on a non-GAAP basis, structurally above foundry margins and below the GPU leader's. As AI mix rises, the trend is up; as leading-edge wafer costs rise, the trend is down; the net is a modest upward drift. Research spending is high and largely expensed rather than capitalized, which is a conservative accounting posture and means GAAP earnings understate the operating reality. Cash flow is healthy โ the light-asset model produces strong free cash flow precisely because capex is minimal โ and the operating-cash-flow-to-net-income ratio sits comfortably above one. The balance sheet is not the problem. The problem is that the growth assumption embedded in the valuation requires the demand curve to do something it has never done at this scale, on a schedule the company cannot fully control.
The key question is not whether AI demand is real. It is real. The question is whether the specific number โ twenty billion by FY28 โ survives the joint probability of all its dependencies. It requires that CoWoS capacity expand on schedule. It requires that Marvell not be crowded out of the allocation queue. It requires that the anchor custom programs ramp on time. It requires that no major customer internalizes a program in the window. It requires that AI capex not pause for macro or power reasons. Each of those is individually likely. The product of all of them is meaningfully less likely than the headline implies.
The Contrarian Cut: What the Headline Hides
Here is where I want to do what I do โ pull the invariant out of the noise and test it. The original report's hidden insight was correct and under-stated: the twenty-billion-dollar target is a demand forecast, not a capacity commitment. I will go one step further. It is a demand forecast with an embedded option on supply, and the option is priced by the market as if it were already exercised. The market is treating a conditional number as an unconditional one. That is the entire mispricing in a sentence.
Consider what a proper valuation would do. It would discount the target by the probability that the external dependencies hold โ CoWoS availability, allocation priority, program schedules, customer behavior. Apply even a modest haircut and the twenty-billion figure becomes an aspirational ceiling rather than an expected value. The distinction matters because a ceiling and an expectation behave completely differently when a single input slips. A ceiling bends. An expectation breaks, and when expectations break, the multiple compresses violently. In the absence of trust, verify everything twice โ and the thing to verify here is whether the number is a promise or a projection. It is a projection wearing the costume of a promise.
The second hidden insight is the nature of the competitive threat. The report named customer concentration as a risk. The real risk is not concentration; it is internalization. Concentration is a revenue mix problem. Internalization is a terminal-value problem. A hyperscaler that builds its own accelerator team does not fire Marvell โ it simply never signs the next contract, and the revenue quietly evaporates with no headline and no churn metric to flag it. This is the quiet killer, and it is invisible in a flash-news framing.
The third hidden insight is the Taiwan tail risk, which neither the headline nor the original report mentioned. It is the largest unhedged exposure in the thesis, and it is unhedgeable by any action Marvell can take. You cannot design your way out of a geographic concentration you did not choose and cannot relocate at will.
The fourth hidden insight is the asymmetry of the two engines. The custom ASIC business is the one that generates the headline growth and the one that carries the internalization risk. The optical DSP business is the one that generates durable, defensible, hard-to-copy revenue and the one the market underprices. If I were allocating against this thesis, I would weight the connectivity engine's durability far more heavily than the market does, because that is where the invariant actually holds. The accelerator engine is optimism. And to borrow a phrase: optimism is a feature, not a bug, until it fails. The custom ASIC bet is optimism about customer behavior, and customer behavior is the least reliable variable in any system.
And there is a fifth, subtler point that connects this whole analysis to my own domain. The reason custom AI silicon exists is the same reason ZK-proof accelerators and decentralized compute networks exist: general-purpose hardware is inefficient for a specific class of workloads, and the economics of scale make specialization worthwhile. The convergence of AI agents and on-chain execution โ the territory I prototyped in 2025, where I built an LLM-driven DeFi executor and hit a wall on the cryptographic signing overhead required to prove agent actions on-chain โ will eventually demand exactly this kind of specialized silicon. The zero-knowledge proof structure I proposed to validate AI decisions without revealing model weights is, at bottom, a call for cheaper specialized compute. Marvell's position in that future is not obvious, but its SerDes and optical interconnect leadership is the connective tissue any such system would need. Code is law until the reentrancy attack โ and the reentrancy attack on the AI-silicon thesis is the moment the demand curve and the supply chain diverge, because that is the seam where the system's assumptions leak.
Takeaway
So what should you actually conclude from a raised number and a rising narrative? Not that AI demand is fake โ it is not. Not that Marvell is a weak company โ it is not; it is the number two custom-silicon house on earth with the number one optical DSP franchise and a genuine design moat. The conclusion is narrower and more useful: the twenty-billion-dollar FY28 figure is a demand ceiling masquerading as a revenue floor, and the entire investment case turns on whether the market correctly prices the conditional nature of that number. If it does, the stock is fairly valued. If it does not โ and the language of the coverage suggests it does not โ then the mispricing is in the gap between the projection and the probability-weighted expectation.
My forward-looking judgment is this. Watch three things, in order of importance. First, CoWoS allocation commentary and any signal that Marvell's packaging priority is slipping relative to the GPU leader โ that is the leading indicator of delivery risk. Second, the customer mix: any hint that a hyperscaler is expanding an internal chip team for a program Marvell currently serves is a terminal-value warning, not a quarterly-miss warning. Third, the optical DSP share, because that is the durable asset, and its stability is the strongest evidence that the underlying franchise is intact regardless of what the headline number does.
The question I would leave on the table is not whether Marvell reaches twenty billion. It is this: when a fabless company publishes a number it cannot unilaterally deliver, whose forecast is it actually publishing โ the company's, or the customer's? Trace the gas trail back far enough and the genesis block is never the designer. It is always the buyer. And the buyer has not signed anything binding. Entropy increases, but the invariant holds โ the invariant being that demand forecasts are not supply guarantees, and the distance between the two is exactly where fortunes are made and lost.