Marvell stock: CyberTAN targets AI networking’s heat problem

CyberTAN unveiled a liquid-cooled AI networking design built around Marvell’s Teralynx T100 specifications on Tuesday, October 13, giving investors a fresh example of how the chipmaker’s technology could reach data centers. The announcement, released at 3 p.m. EDT, describes its Pallas platform as a prototype using a thermal module to simulate the chip’s workload. It discloses no customer order or shipment volume.

For Marvell stock, that makes the development an engineering milestone to examine alongside an already ambitious financial outlook. A cooling system can make a powerful chip usable in a crowded rack. Shareholders ultimately need evidence that customers will qualify the complete equipment, order it and generate profitable sales.

The distinction matters at Marvell’s current scale. A fixed TECHi quote observation showed MRVL at US$265.33 at 3:06 p.m. EDT on October 13, during the Nasdaq regular session, with an equity market value of $232.66 billion. The underlying provider was Yahoo Finance; the feed was marked delayed, with the delay interval unspecified. Those are dated observations, not a live price or evidence that this announcement moved the stock.

What the CyberTAN demonstration establishes

Pallas is designed for the T100’s 102.4-terabit-per-second switching capacity. CyberTAN says its direct-to-chip liquid cooling can handle more than 1,000 watts of localized heat. Its OCP Global Summit exhibit uses a coolant distribution unit and temperature sensors, with a stated 45°C liquid inlet and 40-watt optical modules.

Those specifications describe the intended thermal environment. They do not establish the number of production units a customer has accepted. The practical investor question is whether this equipment can move from a controlled demonstration to repeatable operation in a customer’s own installation.

Think of the development process as a sequence of different proofs. Thermal testing asks whether components stay within acceptable temperatures. Network testing asks whether useful traffic reaches its destination reliably under demanding workloads. Customer qualification asks whether the whole system meets that buyer’s operating requirements. A purchase order and recognized revenue answer different commercial questions again.

An improvement at one stage can be valuable without settling every later stage. That is the appropriate level of confidence for a newly demonstrated architecture.

The chip is established news; the cooling design is new

Marvell introduced the T100 on June 1, saying customer sampling would begin that quarter. Its stated specifications included typical power below 1,000 watts and power consumption up to 25% below competing solutions. These remain company claims, not independently measured results from Tuesday’s event.

The product is a switch chip: it directs data among computing systems. It is not the accelerator that performs the underlying AI calculation. That distinction explains why an investor should look beyond an isolated bandwidth figure. A fast processor still needs a network that delivers the data its workload requires, but networking capacity alone does not measure completed AI work.

Marvell’s July technical explanation describes a single-die design, programmable traffic handling and multiple packaging options. It argues that these features can reduce network complexity and adapt equipment to different configurations. That supplies context for the chip’s appeal; it does not show how a particular customer’s software and traffic will perform on a finished CyberTAN system.

A useful customer test would measure completed work, error rates and energy consumption together. A headline throughput number can look attractive while leaving the buyer’s actual cost per completed job unresolved. Conversely, a modest improvement in system reliability could be commercially useful even without a dramatic change in peak bandwidth.

A 25% chip saving is not a 25% data-center saving

There are two different efficiency claims investors should separate: power used by networking hardware and power used by the facility supporting it.

For the first, consider a deliberately simplified TECHi example. Assume a rack consumes 100 units of power, of which 20 go to the networking equipment being replaced. Assume that entire 20-unit portion becomes 25% more efficient at the same workload. The saving is five units, leaving rack consumption at 95: a 5% reduction for the rack, not 25%.

The 20-unit allocation is an illustrative assumption, not a measured CyberTAN configuration. It is broadly consistent with Marvell’s cited range of 15%–25% for switching and networking’s share of rack power, but the company’s chip-level comparison does not necessarily apply to every component in that range. Optics, other chips, pumps and the remaining system still consume energy.

This is why the relevant question is what the saving covers. If only part of the networking equipment changes, or supporting cooling consumes more power, the net saving will differ. No honest financial model can turn the advertised chip percentage directly into a customer’s total electricity reduction without those inputs.

Even when a power saving is real, a buyer may use the released electrical capacity to install additional equipment. Total consumption could remain unchanged while useful output rises. That would still be valuable, but it would be a capacity benefit rather than a lower utility bill. Investors should ask which outcome the customer actually measures.

The PUE claim needs a facility boundary

CyberTAN also claims a power usage effectiveness, or PUE, of 1.1. The Green Grid defines PUE as total data-center energy divided by the energy used by its IT equipment. It is a facility-efficiency ratio, not a direct measure of a chip’s processing efficiency.

In a simple worked example, 100 units consumed by IT equipment and 110 units consumed by the whole facility produce a PUE of 1.1. The overhead is ten units. That is 10% of the IT load, but about 9.1% of the facility total, calculated as 10 divided by 110.

For a buyer assessing an efficiency claim, the measurement boundary matters as much as the decimal. Does the calculation include the pumps and the equipment rejecting heat outside the rack? Does it reflect a sustained operating period, changing weather and varying utilization? Are both numerator and denominator measured over the same interval?

A demonstration can contribute useful evidence about heat removal. An investor still needs the measurement method before treating its stated ratio as the expected performance of an entire deployed data center. TECHi has not independently measured this system’s efficiency.

Marvell already has a demanding revenue outlook

The latest strategic context is Marvell’s October 6 investor presentation, not an older forecast being presented as current. On slide 34, management raised its fiscal 2028 revenue outlook to $20 billion, from $18 billion. The slide also showed a $12 billion fiscal 2027 outlook.

The revision was 11.1%, calculated as the $2 billion increase divided by the earlier $18 billion forecast. The new fiscal 2028 figure implies 66.7% growth over the $12 billion fiscal 2027 outlook. Both are management estimates, not reported sales.

Using the dated $232.66 billion equity value above, the stock market values Marvell at approximately 11.6 times that $20 billion future annual revenue outlook. This is an equity-value-to-sales illustration. It is not enterprise value, a price-to-earnings ratio or a fair-value target. It holds today’s observed equity value against future revenue and does not adjust for subsequent share issuance, cash or debt changes.

The calculation explains why evidence of execution matters. An investor buying future growth needs to assess both whether the sales arrive and how much profit they produce. A favorable technology demonstration can support confidence in an opportunity without being enough to justify a higher company-wide earnings estimate.

For a deliberately hypothetical downside sensitivity, keep that equity value fixed and reduce the future revenue denominator by 10%, to $18 billion. The ratio rises to 12.9 times. This is not a prediction of a revenue miss or stock price: it simply exposes how much the valuation comparison depends on the sales assumption.

More equipment sales must also produce earnings

Marvell’s second-quarter fiscal 2027 results, for the quarter ended August 1, reported $2.739 billion in revenue, $308 million in GAAP net income and $605.5 million in operating cash flow. Its reported gross margin was 53.1% under GAAP and 58.9% on its adjusted basis.

The 5.8-percentage-point difference between those margins is another reason to state the accounting basis when discussing profitability. Adjusted figures can help explain operating trends, but they do not make the excluded costs disappear. Nor do company-wide margins reveal the profitability of a particular switch design.

A practical model would need the number of chips sold, the revenue Marvell receives per chip and the costs associated with those sales. A system’s complete selling price cannot simply be assigned to its chip supplier. Cooling hardware and other components have their own suppliers and economics.

Readers can use Marvell’s financial statements to follow the reported business as it develops. A prototype announcement belongs beside that evidence, with its narrower scope intact.

The strongest bull case—and the competitive response

The strongest argument for taking this development seriously is that infrastructure constraints can determine whether otherwise attractive chips get deployed. A customer cannot buy an abstract performance specification; it needs working equipment that fits its power, cooling and maintenance arrangements. Progress toward that equipment can reduce a practical obstacle to adoption before it appears in financial results.

There is also fresh evidence that competing platforms are addressing the same obstacle. In a separate 3 p.m. EDT announcement, UfiSpace described a liquid-cooled 102.4-terabit platform using Broadcom’s Tomahawk 6, with 64 ports rated at 1.6 terabits per second. It positioned the design as an early-access architecture for future clusters.

That makes this a competitive implementation problem. A customer’s decision could depend on software, operational experience, support, price and availability as well as thermal design. Two products advertised at the same switching capacity are not automatically interchangeable, and two vendor demonstrations are not a controlled comparison.

The evidence that would strengthen the Marvell case is specific: independently documented performance on real traffic, customer qualification, identifiable production shipments and financial disclosure showing attractive economics. Evidence favoring caution would include deployment delays, a need for costly redesigns or sales growth that fails to improve earnings and cash generation.

For now, Tuesday’s news gives shareholders a concrete system design to track. The next meaningful advance is a demonstrated connection between that engineering work and customer purchases—not another bandwidth figure in isolation.