Two Potential Two Baggers. My Two Largest Positions.
the chat saw these first. and two price targets.
Two stocks. The largest concentrated positions in my account, and not by accident.
The first showed up through Stock Talk Weekly. The second came from Pennycheck. Genuine thanks to both desks. STW called position one early with a clean read on the technical and fundamental setup before consensus had it on the radar. Pennycheck did the kind of primary source verification work on position two that compresses weeks of independent diligence into a weekend of cross checking, and the depth of their documentation is what gave me the foundation to build conviction at size. Neither thesis is mine to claim. The conviction is, and it survived weeks of additional work on both before I sized up.
Position one isn’t a new name on my screen. I’ve followed it for years, traded it in and out across multiple cycles, watched its standards ecosystem inch from PowerPoint slides to one of the largest semiconductor companies in the world publicly backing the standard. Sub-$200M cap connectivity chipmaker. Four design wins on the board with the first start-of-production landing in 2027. An unfolding CFO transition.
Position two is an AI infrastructure systems builder mid-pivot out of legacy memory commodity work into sovereign-grade integration, riding a foundational tri-party alliance the broader market hasn’t fully digested. The piece nobody is modeling is the memory pull-through that runs underneath every server they ship.
Today’s a double feature. Paid only.
Behind the paywall: unit economics, segment math, design win cadence and standards positioning, the read on insider activity and capital structure, catalyst sequencing over the next 6 to 12 months, the disconfirming evidence I’d want to see before I trim, the precise levels where I’m a buyer of weakness, and where I start to ring the register. Two full DCF builds. Four scenarios per name. Position sizing logic and kill switches.
If I’m wrong, you watch the L print in real time. That’s the deal.
If I’m right, this post paid for your subscription twice over before lunch.
Names, theses, sizing, kill switches. All in.
Position one is Valens Semiconductor (NYSE: VLN).
Position two is Penguin Solutions (NASDAQ: PENG).
Both are anchored to the same secular force pulling the next decade of computing through a transition the market is still pricing piecemeal. Massive volumes of high bandwidth data moving from sensors and memory into compute, with the connectivity and integration layers becoming load-bearing infrastructure rather than commodity glue. Position one captures that force at the edge of compute, where sensor data meets the inference brain. Position two captures the same force at the center of compute, where CPU sockets, memory, and AI accelerators converge inside the AI factory.
The underlying force is the same.
Position One: Valens Semiconductor ($VLN)
The setup in one breath
A connectivity chipmaker capitalized at roughly $159M, sitting on $92.6M of cash, no traditional debt, current assets covering current liabilities better than five times over. Trades at 0.95x trailing EV/Sales. Recently announced a 10% headcount reduction with $5M of annualized opex coming out by Q2 close. Standards-level positioning in two distinct ecosystems. Four MIPI A-PHY automotive design wins notched. Samsung publicly backing the A-PHY standard as of September 2025. Q1 print drops May 13.
The market has this priced as a failed automotive SPAC. The reality looks closer to a misvalued connectivity platform with optionality across four different end markets, a balance sheet that takes solvency off the table for at least eighteen months, and a software-grade gross margin profile that survives the revenue trough.
My target inside twelve months is $5 per share. That’s a 114% return from current levels and works out to roughly 3.25x EV/Sales, 3.2x book, a multiple set that the auto and industrial semi peer group routinely supports for businesses with far worse fundamentals than what shows up on Valens’s tape today.
That target is the rerating mark. The actual asymmetry sits past it.
What Valens actually builds
The story everyone reads off a screener is automotive ADAS. That’s wrong, or at least it’s been wrong since management started broadening the cross-industry book three years ago.
Valens designs and sells high performance connectivity chipsets in two segments. Cross Industry Business covers audio video, video conferencing, industrial, medical imaging, and digital signage. Automotive covers ADAS and autonomous driving sensor connectivity. The technical product set spans the VS, VA6000, and VA7000 families, anchored by two standards the company holds material IP on: HDBaseT and MIPI A-PHY.
HDBaseT is the older of the two and lives mostly in the cross industry segment. It moves uncompressed ultra high definition video, Ethernet, USB, control signals, and power across long copper runs through a single cable. Pro AV installers, hospital imaging suites, classroom systems, and corporate video conferencing rooms all use it. The TAM doesn’t grow heroically, but it doesn’t roll over either, and the installed base produces a steady refresh cycle that funds ongoing R&D without putting the cross industry segment at the mercy of any single customer.
MIPI A-PHY is the newer angle and the one that matters for the rerating thesis. A-PHY is an industry standards body specification for long reach, high speed serializer/deserializer connectivity originally scoped for automotive ADAS and autonomous driving camera and sensor links. Valens shipped the world’s first A-PHY compliant chipsets in the VA7000 family. As of CES 2026, the company has notched four global design wins for A-PHY chipsets, the most recent landing with a premium automotive OEM serving the Chinese market with a 2027 start of production timeline.
Samsung announced support for MIPI A-PHY as a high speed sensor connectivity standard in September 2025. That detail matters more than the analyst community has metabolized. When the largest mobile imaging supplier in the world endorses a connectivity standard, the standard’s path to ubiquity flips from possibility to gravitational. Standards adoption tends to look slow until it doesn’t. Then it looks like Ethernet.
Why the market mispriced this
Three reasons, in descending order of relevance.
First, the post-SPAC overhang. Valens went public via SPAC in September 2021 at a deal value north of $1.1B. The market cap today is one seventh of that. Anyone who bought near the deal price has spent four years carrying a position that did nothing but bleed. The selling pressure from those holders is a real flow phenomenon and it’s largely exhausted. Insider Form 4 filings over the trailing 90 days show routine 10b5-1 dispositions, not capitulation. The natural sellers have sold.
Second, screener exclusion. At $159M market cap, Valens sits below the threshold most institutional auto and industrial semi screeners use. Sell side coverage is thin: seven analyst tags, only four submitting active estimates per the most recent aggregator data. The name is not on most generalist tech buy side desks because the format and the size and the sector are not on the radar. That orphan coverage profile is a feature, not a bug, when you are a research aggregator. Crowded names trade at fair value. Orphans don’t.
Third, the segment optics. Public reporting still bins this as Cross Industry plus Automotive, and the automotive line has been ugly through the OEM ADAS digestion cycle. What the segment optics obscure is the diversification underneath. The VA7000 IP block is engineered for automotive bandwidth profiles, but the technical envelope, long reach, low latency, high throughput sensor data transport, applies one for one to industrial machine vision, medical imaging, video conferencing, and the emerging robotics edge AI build out. The work to repackage that IP into adjacent verticals is already underway. The MCNEX joint product launch on QHD camera modules through unshielded twisted pair, announced in early 2026, is a forward indicator of what that diversification looks like at the product level.
The robotics and edge AI optionality
Robotics is moving from simple programmed automation toward perception heavy autonomy. That transition requires more cameras and more sensors per unit, all feeding inference models that themselves need to receive data with low latency and high reliability. The bandwidth between the sensor head and the compute brain is the unsexy bottleneck nobody talks about.
A-PHY was designed for automotive sensor links. Industrial robotics, autonomous mobile robots, medical surgical platforms, and edge AI camera systems share the technical requirement profile. Long copper runs, electromagnetic interference resilience, deterministic latency, embedded power. Valens already has the silicon and the standards seat. The work to convert that into design wins outside automotive is a sales cycle problem, not a technology problem.
Place Valens in the edge AI plus robotics basket alongside Synaptics, One Stop Systems, Vishay Precision Group, and CTS. The four cover different layers of the same architecture. SYNA sits closer to the edge processor and human machine interface side. OSS handles rugged edge compute. VPG and CTS cover physical world sensing, measurement, and actuation. Valens slots in as the high bandwidth nervous system between sensor and compute. The basket is complementary, not redundant, and Valens carries the lowest valuation profile of the group.
The financial profile
Trailing twelve month revenue lands at $70.6M. Gross margin runs 62.4% on a GAAP basis and 66.7% on a non-GAAP basis per recent quarterly disclosures. Operating margin of negative 48.2% reflects the cost structure carry over from the SPAC era and the ongoing investment in A-PHY ecosystem development. Operating cash flow ran negative $12.7M for FY25.
Management guides 2026 revenue to $75M to $77M. Q1 2026 guide is $16.3M to $16.7M. The Q1 print on May 13 will tell us whether the cross industry recovery management has been signaling for two quarters is actually showing up in the numbers.
The balance sheet is where this stock stops looking like a small cap money pit and starts looking like a misvalued option. Current cash and short term deposits of $92.6M against current liabilities of $22.9M and current assets of $118.7M. No traditional debt, no maturity wall, total shareholders equity of $105M. At the current burn rate, runway extends past 36 months even before the announced opex cuts begin to take effect.
Solvency is not the question here
The 10% headcount reduction announced January 2026 is expected to deliver $5M in annualized opex savings by Q2 close. That’s roughly a 13% reduction in the operating loss before any revenue contribution from new design wins. The implied 2026 operating cash burn, after the cost actions land, sits closer to negative $7M to negative $9M.
Cash position covers that burn for over a decade at the post-restructuring rate. There is no plausible path to insolvency inside any reasonable thesis horizon, and no near term need for dilutive capital. That fundamentally changes the risk profile. The downside scenario for Valens is multiple compression and time decay, not zero. Most small cap semiconductors carrying a similar revenue profile cannot make that claim.
Comp set and target multiple math
The auto and industrial connectivity peer group, screened on EV/Sales (NTM) and gross margin similarity, supports a 3x to 5x EV/Sales multiple in normalized conditions. Valens currently trades at roughly 0.88x EV/Sales on the 2026 guide midpoint. A move to 3.0x EV/Sales on the 2026 guide implies an enterprise value of approximately $228M, which, layered on top of the $92.6M cash position, produces an equity value of $321M, or roughly $4.72 per share at the current share count.
Push the multiple to 3.5x on the strength of A-PHY design win cadence and the cross industry recovery, and the math walks to $5.30 per share. The target $5 mark is therefore not aspirational. It’s the midpoint of a reasonable peer set rerating on the existing revenue base. Anything beyond that requires the operating leverage to start showing, which is the Bull and Blue Sky territory in the scenarios below.
Discounted cash flow
I built a five year forecast through 2030 with explicit revenue, gross margin, operating margin, and operating cash flow assumptions. Discount rate of 12%, terminal growth of 2.5%. Net cash position layered onto the present value of free cash flows.
Base case DCF. Revenue grows at 10% CAGR from $76M (2026E) to $114M (2030E). Gross margin holds at 62%. Operating margin recovers from negative 25% in 2026 to positive 11% in 2030 as the opex base flexes against revenue growth. Operating cash flow walks from negative $15M to positive $12M across the five years. Sum of present values of OCFs runs negative $11M. Terminal value contributes approximately $73M of present value. Adding back the $92.6M net cash position produces an equity value of $154M, or $2.27 per share. The base case effectively supports the current price. That’s the floor.
Bull case DCF. Revenue compounds at 20% CAGR to $157M by 2030, driven by A-PHY automotive ramps starting in 2027 plus traction in adjacent verticals. Operating margin reaches breakeven by 2028 and positive 22% by 2030. OCF walks from negative $10M to positive $35M. Sum of present values of OCFs lands at approximately $27M. Terminal value contributes $214M of present value. Equity value of $334M produces $4.92 per share.
Blue Sky DCF. Revenue compounds at 27% CAGR to $220M by 2030, reflecting A-PHY adoption acceleration, named robotics customer wins, and operating leverage flowing through. Operating margin reaches positive 36% by 2030. OCF walks from negative $8M to positive $80M. Sum of present values of OCFs runs $92M. Terminal value contributes $490M of present value. Equity value of $675M produces $9.94 per share.
Bear case DCF. Revenue stagnates near $70M for three years, then resumes modest growth. Cash burn outpaces the announced opex cuts. Management raises capital in 2027 at a discount. Diluted share count rises 10%. Equity value compresses to $55M, producing roughly $0.75 per share post dilution.
The DCF tells the asymmetry story cleanly. Base case supports the current price. Bull case delivers the double. Blue Sky delivers four to five times. Bear case produces a 70% loss with capital impairment. Probability weighted, with my own subjective weights of 35% bull, 35% base, 15% blue sky, 15% bear, the implied fair value sits at approximately $4.10 per share, or roughly 75% above current.
Technical setup
The stock has been pinned right at the prior high of $2.48 across multiple recent sessions. On the daily, the 9 and 21 EMA cluster has been reclaimed and the stock is tucked under the 200 day. On the weekly, the 9 and 21 EMAs have been reclaimed and the stock sits below the 50 week. This is the moving average pinch, the setup where multiple timeframes are compressing into a single decision point.
A monthly close above the 9 and 21 EMA cluster on the monthly chart would be the first such close in the company’s public history. The structural breakout potential is real. I am not running the position on technicals alone, but the chart aligns with the fundamental setup, which is the timing edge worth noting.
Catalysts on the calendar
The Q1 print on May 13 is the proximate catalyst. The CFO transition resolution, with Nathanzon’s notice already filed, is the personnel catalyst. The fifth A-PHY design win, if announced inside 2026, is the standards momentum catalyst. Any signed customer outside automotive, in industrial, medical, or robotics, is the diversification catalyst that triggers the multi vertical rerating. Any of these can reprice the stock 30 to 50% on its own, and combinations of them produce the path to the $5 target without requiring a multiple expansion beyond the peer group’s normalized range.
Risks and what would invalidate the thesis
Cash burn outrunning the announced opex cuts. If the Q2 print shows operating loss not improving meaningfully against the $5M annualized savings figure, the runway math changes and dilution risk reenters the picture.
A-PHY ecosystem adoption stalling. If no new design wins land inside 2026 and the existing four don’t show signs of expanding scope, the standards moat thesis weakens and the stock reverts to trading on automotive segment optics.
Cross industry recovery failing to show. Management has been signaling cross industry recovery for two quarters. If Q1 disappoints and Q2 guidance walks lower, the revenue trajectory the bull case requires is gone.
CFO transition disorder. The departure of a CFO at a small cap with negative cash flow is a valid yellow flag. If the replacement search drags or signals deeper governance issues, that’s a separate problem that needs reassessment.
Take out at the wrong price. At a $159M market cap with $92.6M cash and standards level IP, Valens is a credible take out target for a strategic buyer. A take out anywhere below $4 per share would underperform the bull case despite locking in a positive return. The risk of a near term low ball offer is real, although a board with the Series B and SPAC sponsor history Valens carries is unlikely to accept anything below book value plus a meaningful premium.
Position Two: Penguin Solutions ($PENG)
The setup in one breath
An AI infrastructure systems builder formerly known as SMART Global Holdings, repositioning into pure play AI factory architecture and integration. Trading at a forward P/E of approximately 8 against a 2026 outlook that was raised in March. Three concurrent strategic alignments running underneath the headline story: a 13 year Intel CPU partnership that anchors federal HPC and nuclear weapons lab deployments, a parallel AMD partnership delivering production scale enterprise clusters, and a tri party alliance with SK Telecom and SK hynix that is now generating disclosed related party revenue inside the income statement. Add a free option on photonic memory through a documented multi year investment in Celestial AI that monetized into Marvell shares plus a contingent earnout participation worth up to $2.25B.
Most sell side analysts are modeling Advanced Computing as the lumpy revenue line, Memory as the steady grower, and Optimized LED as the legacy drag. That framework misses the load bearing wall. Every server Penguin ships pulls 4 to 8TB of SMART Modular memory content. The memory pull through is structurally tied to the server pipeline, which is structurally tied to the Intel and AMD CPU cycles, which are structurally tied to the AI capex super cycle. The unit economics flow downhill. Most models do not reflect that.
My twelve month target on PENG is $52 per share, roughly double from current levels. The path to that target runs through the memory cycle, the SKT preferred conversion math, and the PMA optionality, with the Celestial earnout sitting on top as a tail event that nobody is paying for today.
The 13 year Intel partnership
Penguin Computing has been Intel’s named HPC server partner since 2013. Intel publicly named the company its “HPC Technical Computing Partner of the Year” at the 2016 Intel Solutions Summit. Three named Intel VPs publicly endorsed the partnership across multiple product cycles in the years that followed: Phil Harris, Trish Damkroger, and Hugo Saleh.
The federal credentials are where the partnership shows up as contractual reality rather than narrative. Penguin holds active deployments at all three United States nuclear weapons national laboratories, anchored by the $39M NNSA CTS-1 contract that delivered 22+ petaflops of Intel Xeon powered systems to Los Alamos, Sandia, and Lawrence Livermore. The Magma supercomputer at LLNL, deployed under that contract, ran Intel Xeon Cascade Lake Platinum 9200 series processors. Per the LLNL announcement: “Magma is one of the first deployments of Intel’s Xeon Cascade Lake Platinum 9200 series processors, which are specifically designed for high performance computing machines.”
The CTS-2 procurement cycle is the forward catalyst. Penguin sits on the credentialed bidder list for the successor contracts, and the federal supplier moat is one of the cleanest you can find in semis adjacent businesses. Security clearance, decades of delivery track record, established supplier relationships. New entrants do not show up overnight.
The Relion server line transitioned fully to Intel Xeon Scalable in 2017, and as of May 2026 the live product catalog spans 17 Intel based SKUs across 1U through 8U and OCP form factors. The flagship AI servers shipping today, the XE4418GTS-DTC and XE8418GTS, pair Intel 6th Gen Xeon Scalable with NVIDIA B300 HGX GPUs, configured with 8TB DDR5 at 6400 MT/s and MRDIMM modules running up to 8000 MT/s across 32 DIMM slots.
The memory load bearing wall
This is the part the sell side is underweighting.
Every Intel Xeon 6 server Penguin ships pulls 4 to 8TB of DDR5 and MRDIMM content. SMART Modular, a Penguin product brand, makes those DIMMs. The same revenue lands on Penguin’s income statement twice: once on the server itself, once on the memory inside the server. That’s not an accounting trick. It’s a vertical integration capture on the highest dollar value component in the AI server BOM.
The Intel Xeon 6900 series with 12 DDR5 channels and MRDIMM at 8800 MT/s aligns directly with management’s stated memory growth thesis. Q2 FY26 memory segment revenue grew 63% year over year to $172M. The pull through is not theoretical. It is showing up in the printed numbers, and the Venice 2026 step up extends the runway.
For AMD Venice EPYC servers in 2026, the configuration math gets larger: 16 DDR5-12800 channels times 2 sockets equals 3 to 4TB per server versus 1.5TB on Turin. Higher channel counts, higher per channel speeds, higher memory content per server. The memory pull through compounds with each CPU generation upgrade because the server platform demands more memory bandwidth, not less.
Strip out the photonics excitement, the SKT alliance, and the earnout optionality, and the load bearing wall is just this. CPU pipelines drive server pipelines. Server pipelines drive memory pipelines. Memory grew 63% year over year last quarter and the next two CPU generations both raise memory content per server. The base case for PENG is that growth print, repeated.
The AMD parallel partnership
The market reads PENG as Intel exposure. The reality is dual vendor.
Penguin management named AMD as a core partner on the Q1 FY25 earnings call alongside NVIDIA, CDW, and SKT. From the call summary: “Launched a rapid development workshop program and continued partnerships with NVIDIA, AMD, CDW and SKT; development work on ICE ClusterWare and optical memory appliance (OMA) underway.”
The partnership precedes the AI boom by years. Penguin’s CTO Phil Pokorny on the Penguin/AMD/Shell case study: “AMD partnered with Penguin Solutions years ago to achieve early access to new technologies, around the same time that Shell IT first leveraged AMD EPYC processors in their own technology. Since they had history, a natural partnership was forged.”
The execution proof point is the Shell HPC cluster running in a Houston data center: 864 dual socket Penguin Altus servers powered by 96 core 4th Gen AMD EPYC 9654 processors, totaling 1,728 processors and 165,888 cores, deployed with immersion cooling to manage 700W plus TDPs. It is the flagship reference design Penguin and AMD co market at industry events.
Why this matters for the thesis. Customer chooses CPU. Penguin captures the integration revenue and the memory revenue regardless of which CPU vendor wins the socket. AMD’s server CPU TAM is on track to roughly double to $120B by 2030 on credible analyst forecasts. Penguin’s exposure to that TAM growth runs through the integration and memory layers, which are CPU agnostic by design. The dual vendor architecture is a competitive moat the sell side does not adequately credit.
The SK Telecom alliance and the Korean sovereign AI stack
December 13, 2024, SK Telecom invested $200M in Penguin Solutions through SPV Astra AI Infra LLC. The instrument was 200,000 convertible preferred shares at $1,000 per share. Conversion price into PENG common is $32.80784. Dividend of 6% per annum, cumulative, payable quarterly in kind or in cash at PENG’s option. Auto conversion mechanic: after year two, if PENG common VWAP exceeds 150% of conversion price for 15 consecutive trading days, the preferred auto converts. That trigger price works out to $49.21 per share. Min Yong Ha from SKT joined the PENG board.
The capital was not passive. At CES 2025, PENG, SK Telecom, and SK hynix signed a three way strategic collaboration agreement to develop and deliver comprehensive AI data center solutions branded AIDC. From the SK hynix corporate statement issued January 2025: “Innovation of semiconductor memory technology is essential for the efficient operation of AI data centers, with improving power efficiency and heat dissipation performance being a key challenge. Through collaboration with SK Telecom and Penguin Solutions, we plan to overcome the limitations of memory technology and lead the expansion of the global AI ecosystem.”
The execution proof point is the Haein cluster at SKT’s Gasan AI Data Center in Seoul. One of South Korea’s largest GPU clusters: 1,000+ NVIDIA Blackwell GPUs integrated into a single supercomputer, end to end designed, built, and deployed by Penguin Solutions with a Korea based execution team. Production ready August 2025 within weeks of all materials arriving onsite. Selected by Korea’s Ministry of Science and ICT as part of the Proprietary AI Foundation Model Project supporting Korean language LLMs and sovereign AI workloads. Successor cluster to SKT’s H100 based GPUaaS launched December 2024.
The financial implications are now visible in the filings. Q2 FY26 10-Q discloses related party revenue from SK Telecom’s affiliate of $0.9M for the quarter and $33.1M for the six months ended February 27, 2026. PENG paid $6.1M in preferred stock cash dividends in 1H FY26 tied to the SKT preferred shares, or $3.0M per quarter. SK hynix publicly grouped Penguin Solutions with TSMC and NVIDIA as core ecosystem partners for AI data center memory commercialization in October 2025.
The conversion math is worth running. The preferred is convertible into approximately 6.10M shares of PENG common at the $32.81 conversion price. That’s roughly an 11% theoretical dilution against the current share count. The auto conversion trigger requires PENG to trade above $49.21 for 15 consecutive sessions. If the stock reaches that level, the dilution is offset many times over by the appreciation of the underlying common. The preferred is structured such that it costs Penguin shareholders nothing if the thesis works and a manageable dividend cost if it does not.
Photonic memory: Celestial, Marvell, and the PMA
Penguin Solutions, then SMART Global Holdings, participated as a named investor in Celestial AI’s Series B in June 2023 alongside Samsung Catalyst, Koch Disruptive Technologies, Temasek’s Xora, Porsche, and IAG Capital. The Series C1 announcement in March 2025 publicly listed Penguin as a strategic supporter alongside AIN Ventures, IMEC XPand, and Samsung Catalyst.
December 2, 2025, Marvell announced the acquisition of Celestial AI for $3.25B upfront ($1.0B cash plus $2.25B Marvell stock) plus up to $2.25B contingent earnout if Celestial cumulative revenue exceeds $2B by end of Marvell FY29. Closing date February 2, 2026.
Penguin received approximately $32M cash plus Marvell shares as proceeds from the disposition of its Celestial position and recognized a $27.5M GAAP gain in Q2 FY26. Per CFO Nate Olmstead on the Q2 FY26 earnings call, April 1, 2026: “Approximately $32 million received from proceeds from the disposition of our investment in Celestial AI in connection with its sale to Marvell Technology.”
What got buried inside that disclosure is the participation in the contingent earnout. Penguin retains a pro rata interest in the $2.25B contingent consideration if Celestial cumulative revenue exceeds $2B by the end of Marvell FY29. Per ASC 815 scope exception, Penguin will recognize earnout gains only when realizable, which means as future non operating income surprises. Marvell management has guided Celestial revenue trajectory to $500M annualized run rate by Q4 FY28 and $1B annualized run rate by Q4 FY29. If those guides hit, the earnout is in the money. The street is not modeling that.
The Photonic Memory Appliance
The other half of the Marvell partnership is the on going operational alliance, not the financial earnout. Penguin Computing is the named chassis builder for the Photonic Fabric Appliance, a 2RU rack mount system holding 16 of Celestial’s Photonic Fabric Modules. PFA delivers 33TB unified shared memory and 115 Tbps all to all switching across 16 connected XPUs per the technical disclosures in Celestial’s Hot Chips 2025 presentation and arXiv paper 2507.14000.
Penguin’s branded variant is the Photonic Memory Appliance, formerly named OMA, integrating Celestial’s photonic silicon with SMART Modular memory expertise. Initial PMA shipments are targeted for late 2026 to early 2027.
PENG CEO Kash Shaikh referenced the PMA, the Celestial investment, and the active Marvell partnership in three separate prepared statements and the most consequential Q&A exchange of the Q2 FY26 earnings call. From the prepared remarks: “In parallel, we continue to advance development of our Photonic memory appliance or PMA, formerly referred to as OMA, which is designed to extend memory capacity and bandwidth for large scale AI environments.”
When Loop Capital’s Ananda Baruah asked directly whether photonics is required for CXL scale up, Shaikh’s response confirmed the architecture and the partnership status: “In terms of photonic memory appliance that we are working on in our partnership with Celestial AI, which is now obviously Marvell, that provides increased capability because obviously, when you have photonic connectivity, then you have increased capacity to share the memory. So it takes it to the next level. However, CXL in itself is an advantage. We can take it to the next level with the photonic appliance.”
Translation. Penguin’s near term memory thesis runs on CXL based architectures. Photonic memory through the PMA is the next layer of memory scale, which Penguin builds, brands, and ships. Marvell now owns the underlying Celestial silicon. The integration partnership remains live. The earnout sits on top as a financial upside on Celestial’s revenue ramp.
PMA architectural advantages
Per Celestial’s technical disclosures, the PMA architecture delivers material specification advantages over current GPU memory architectures. Optical bandwidth per Photonic Fabric Module of 7.2 Tbps each direction. Aggregate switching of 115 Tbps all to all per appliance with radix of 16. Unified memory address space up to 32TB across 16 XPUs or servers. Module composition built on TSMC 5nm ASIC including an 8 Tbps switch, 2 HBM3e controllers, 4 DDR5 controllers, 2 HBM3e stacks, photonic interposer, and fiber array unit in 2.5D package. Latency to remote memory via PCIe attached PF NIC under 250 nanoseconds. Energy efficiency of approximately 6.2 picojoules per bit versus approximately 62.5 picojoules per bit on NVLink and NVSwitch, an order of magnitude advantage.
For inference workloads in particular, where memory bandwidth bottlenecks the throughput far more than compute does, an architecture that adds 33TB of pooled memory at near nanosecond latency to a 16 XPU cluster fundamentally changes the unit economics of large scale serving infrastructure. PMA is not a marginal product. It’s a category creating product, assuming the engineering execution lands inside the disclosed timeline.
Discounted cash flow and target
I’m not building a full multi year DCF for PENG in this letter the way I did for VLN, because the Advanced Computing revenue lumpiness makes any single point estimate misleading. Instead, I’m using a forward EBITDA framework anchored to consensus 2027 estimates and cross checked against the segment math.
Consensus 2027 EBITDA of approximately $310M against an enterprise value of approximately $1.45B implies 4.7x EV/EBITDA on the forward year. Peer set comparables for AI infrastructure systems integrators trade in a 7x to 10x EV/EBITDA range. Applying 8x to the 2027 EBITDA estimate produces an enterprise value of $2.48B. Adjusted for net cash and the SKT preferred (treated as debt at conversion), the implied equity value walks to approximately $52 per share, or roughly double from current.
That target does not include the Celestial earnout participation, which is structurally unmodeled. It does not include any contribution from PMA revenue, which begins to flow in late 2026. And it does not include the optionality on additional sovereign AI engagements outside Korea. Each of those is a separate call option layered onto the base thesis.
Scenarios
Blue Sky. PMA ships on schedule with initial shipments late 2026 or early 2027. Celestial cumulative revenue exceeds $2B by Marvell FY29, activating PENG’s pro rata claim on the $2.25B earnout. Memory segment compounds off the Q2 base of $172M growing 63% year over year. Korea sovereign AI stack expands beyond Haein into additional clusters. Penguin becomes the named integrator for photonic memory at scale. SKT preferred auto converts above $49.21. Stock trades through $75.
Bull. Memory cycle delivers per the 65 to 75% segment guide. Venice 2026 step up materializes. Marvell hits its own Celestial trajectory of $500M ARR Q4 FY28 and $1B Q4 FY29. PMA validates with named design wins. SKT preferred auto converts. Stock walks to $52 to $58 inside twelve months.
Base. Memory grows in guide. Advanced Computing remains lumpy. PMA slips past the initial window with modest first year revenue. Celestial earnout treated as a free option with no street modeling. Optimized LED drag persists. Multiple compresses as AI capex narrative normalizes. Stock holds between $30 and $36 with positive operating cash flow conversion.
Bear. Hyperscaler capex digestion in the second half of 2026. Memory pricing rolls over. Marvell post acquisition routes Celestial silicon directly to hyperscalers, compressing the integrator layer margin. Celestial misses the $2B threshold and the earnout zeros. Customer concentration bites simultaneously with Shell completing its build out, SKT decelerating, and the NNSA contract gap widening. SKT preferred dilution lands without offsetting common appreciation. Stock revisits its 2022 to 2025 trading range of $14 to $22.
The probability weighting I’m running, with my own subjective weights of 30% bull, 35% base, 15% blue sky, 20% bear, produces an implied fair value of approximately $42 per share, roughly 60% above current.
Risks
Hyperscaler concentration in Advanced Computing. The segment has historically been one or two customer dependent. A pause in any single customer’s capex schedule produces lumpy quarterly prints that the market punishes regardless of underlying thesis health.
Memory commodity exposure. Memory is a cyclical product with pricing volatility. The 63% year over year growth rate is not extrapolable. The cycle turns. When it turns, the memory segment compresses revenue and gross margin simultaneously, which compounds the downside.
PMA execution risk. Initial shipments late 2026 or early 2027 is a tight timeline for a product category that does not yet exist commercially. Engineering slippage of two to three quarters is plausible. Slippage of more than that pushes the PMA optionality past the thesis horizon.
Celestial earnout dependency. The earnout is contingent on Celestial cumulative revenue exceeding $2B by end of Marvell FY29. Marvell is guiding the trajectory aggressively. The reality is Marvell now owns Celestial and may rationally choose to integrate the silicon vertically rather than commercialize externally at the pace the earnout requires. The earnout is a real option, not a guaranteed payout.
Federal contract recompete risk. The CTS-1 contract anchors the federal credentialing story. If the CTS-2 successor procurement lands with a different prime, the federal anchor weakens materially. The credentialing moat is real, but federal contract continuity is never automatic.
SKT preferred overhang. The 6% dividend obligation runs at $12M per year. If the stock fails to clear the auto conversion trigger, the dividend is a perpetual cash drag. The preferred is only positive optionality if the stock reaches $49.21.
Why these two together
Both names sit in semiconductor adjacent territory. Both names are mispriced for different reasons. Both names carry asymmetric upside profiles. The correlations are not what they look like on the surface.
VLN exposure is to the edge of compute. Sensor data moving from physical world inputs into edge AI inference systems. The end markets are robotics, industrial automation, automotive, medical imaging, video conferencing. The macro driver is the build out of inference at the edge over the next decade.
PENG exposure is to the center of compute. AI factory architecture, sovereign AI buildouts, federal HPC, and the memory pull through that runs underneath every CPU socket the company ships. The end markets are hyperscalers, enterprises, sovereign nations, and federal agencies. The macro driver is the AI capex super cycle and the secular shift toward in country compute capacity.
In a hard hyperscaler capex digestion scenario, PENG compresses. VLN does not, because VLN’s revenue base is not levered to hyperscaler capex. In a federal procurement gap scenario, PENG compresses on the NNSA recompete. VLN does not. In an A-PHY ecosystem stall, VLN compresses. PENG does not.
The two positions hedge each other across non market specific risk vectors while sharing exposure to the same long term secular force. That is a feature of the construction, not a coincidence. I sized both up with full awareness of the correlation structure. The convexity comes from owning the exposure at small cap multiples on both ends of the compute infrastructure stack, while most of the market is paying mid cap multiples for the middle of the stack.
Catalyst calendar through Q4 2026
May 13, 2026: VLN Q1 print. The proximate catalyst on position one. Cross industry recovery commentary, A-PHY design win update, and any pre announcement of additional ecosystem signal.
Late May to June 2026: VLN CFO replacement announcement. Resolution of the personnel transition triggered by Nathanzon’s notice.
July 2026: PENG Q3 FY26 print. Memory segment growth rate, Advanced Computing revenue trajectory, PMA development update, and any incremental commentary on Korean sovereign AI engagement expansion.
H2 2026: PMA initial shipment ramp. The proximate optionality catalyst on position two. Any named customer announcement, pricing benchmark disclosure, or production validation moves the stock 10 to 20% on its own.
Q3 to Q4 2026: VLN A-PHY design win five and beyond. The standards momentum catalyst on position one. Each additional win, particularly outside automotive in robotics, industrial, or medical, recalibrates the diversification thesis.
Q4 2026 to Q1 2027: PENG SKT preferred auto conversion trigger. Requires PENG to trade above $49.21 for 15 consecutive sessions. If reached, the conversion absorbs the dilution overhang and removes the dividend cost.
Q4 2026: Marvell Celestial revenue trajectory commentary. The first marker on whether the Celestial earnout is tracking toward the $2B cumulative threshold by FY29.
Throughout 2026: NNSA CTS-2 procurement cycle developments. The federal anchor recompete is the longest dated and least telegraphed catalyst on the calendar.
Closing
VLN is balance sheet resilient, standards moat protected, optionality rich, and trading below cash burn coverage on a multi year basis. The market is pricing it as a failed SPAC. The reality is a mispriced connectivity platform with edge AI and robotics optionality the auto segment optics obscure. Target $5 inside twelve months. Path to $10 inside three years if the design win cadence continues to compound.
PENG is the AI infrastructure systems integrator carrying three concurrent strategic alignments underneath an 8x forward earnings multiple. The load bearing wall is memory pull through, layered with CXL based architecture in the near term, the PMA platform beyond that, and the Celestial earnout sitting on top as an unmodeled tail. Target $52 inside twelve months. Path to $75 plus if the sovereign AI expansion compounds and PMA validates.
Both positions are sized to matter, both have explicit kill switches, and both print results inside the model portfolio in real time. Win or lose, you watch it.
The chat sees the trades first.
Names, theses, sizing, kill switches. All in.