PENG isn't a server company anymore. It's a CXL memory pure-play — and the breakout just got confirmed.
Memory is now 50% of revenue and grew 63% YoY in Q2 FY26. Advanced Computing dropped 42% — on purpose. The May 5 after-hours $40 print was real: PENG hit a fresh 52-week high of $39.93 on May 7 and is consolidating $39+ in cash hours. Marvell's $3.25B Celestial AI acquisition closed February 2, 2026, with PENG as the photonic memory appliance partner. At ~1.45x sales, the rerate has barely started. The breakout is now confirmed, not anticipated.
This is a memory company in a server company's wrapper. The market just figured it out.
Penguin Solutions used to be SMART Global Holdings. Same business, new name (October 2024), new CEO (Kash Shaikh, his first earnings call was Q2 FY26 on April 1, 2026). Three segments on paper: Advanced Computing, Integrated Memory, Optimized LED. In reality, the only segment that matters now is Memory — it's 50% of revenue, growing 63% YoY, and it's where the entire AI thesis lives. Advanced Computing dropped 42% YoY in Q2 because management is deliberately exiting hyperscaler concentration and winding down the Penguin Edge legacy line. The CEO is essentially burning down the old business to fund the AI pivot. Q2 numbers showed the strategy is working at the unit economics level. Bottom line was a +41% beat. The May 5 after-hours $40 print was the leading indicator. The May 7 cash session confirmed it — fresh 52-week high of $39.93, closing $39.72. The breakout is real and the rerate is happening in real time. At ~1.45x sales, this is still cheap relative to where AI infrastructure pure-plays trade.
The hidden layer that solves the GPU memory wall.
If you do not understand CXL memory, you cannot value PENG at $40. This is the technology underneath the entire pivot. Without it, MemoryAI KV Cache is just another server box. With it, PENG owns a piece of the inference economics that hyperscalers and neoclouds are converging on.
A modern NVIDIA H100 GPU has 80GB of HBM (High Bandwidth Memory) on-package. To run large language models with long context windows and high concurrency, inference workloads need terabytes of memory. The current solution: buy more GPUs, most of which sit idle on compute but are bought for their memory. That is operationally wasteful at $30K+ per GPU.
The Core Architectural Shift
Why The Inference Shift Drives This
Through 2024, AI was overwhelmingly about training — running large models on massive GPU clusters for weeks at a time. Training workloads are memory-intensive but in a different pattern (high bandwidth, less random access). Through 2025-2026, the workload mix is shifting decisively toward inference — actually serving end-user queries at scale. Inference workloads need:
| Inference Requirement | What It Means | Why CXL Helps |
|---|---|---|
| Long context windows | 200K+ token prompts (legal docs, codebases) | KV cache size grows linearly — needs pooled memory |
| High concurrency | Thousands of simultaneous user queries | Each query holds KV state — pooled memory enables sharing |
| Low latency | Sub-500ms time-to-first-token | CXL.mem latency is sub-100ns — fast enough to feel local |
| Cost per token | Inference economics drive deployment scale | Memory at fraction of GPU cost = lower capex per token served |
What MemoryAI KV Cache Actually Does
PENG's MemoryAI line includes both CXL-based servers and the dedicated MemoryAI KV Cache Server. The KV cache is the technical innovation: every LLM query stores intermediate "key-value" pairs that get reused across token generation. If those caches live in pooled CXL memory instead of GPU memory, the GPU is freed up for compute, not memory storage. The result is higher GPU utilization, lower per-query memory cost, and faster time-to-first-token.
The Tier 1 financial institution win disclosed on the Q2 FY26 call validated this — they deployed CXL-powered KV Cache servers for an on-premise AI factory, exactly the use case management has been pitching. CFO Nate Olmstead said on the call: "CXL adoption is timely given the transition to inference because... you need increased memory for faster LLM responses."
Celestial AI is now Marvell. PENG kept the partnership.
The longer-term thesis is the Photonic Memory Appliance (PMA), formerly called OMA (Optical Memory Appliance). PENG was an early strategic investor in Celestial AI alongside AMD Ventures, Samsung Catalyst Fund, Fidelity, BlackRock, Tiger Global, and Lip-Bu Tan. Celestial built optical interconnect technology — light-based data movement instead of copper — that solves the next bottleneck after CXL: getting memory to scale beyond a single rack.
The Acquisition Timeline
| Date | Event | Detail |
|---|---|---|
| Aug 13, 2025 | Celestial AI Series C1 close | $255M raise, $520M total funding. PENG already strategic investor. |
| Dec 2, 2025 | Marvell announces Celestial AI deal | $3.25B base price ($1B cash + 27.2M MRVL shares), up to $5.5B total with earnouts. |
| Feb 2, 2026 | Acquisition closes | Marvell completes deal. Celestial AI CEO David Lazovsky joins MRVL. |
| Apr 1, 2026 | PENG Q2 FY26 earnings | PENG discloses ~$32M monetization of Celestial stake, $27M GAAP gain. Photonic partnership continues. |
| FY2028 (target) | Marvell-stated revenue accretion | Marvell CFO: Celestial expected to be accretive to non-GAAP earnings once meaningful revenue begins H2 FY28. First earnout ($500M cumulative Celestial revenue by FY29). |
What This Means For PENG Specifically
The PMA chassis is built by Penguin Computing using Celestial's photonic devices. The architecture combines 48-72GB HBM3e and 2TB of DDR5 memory with 7.2 Tbps bandwidth in each direction per module. Sixteen modules in a 2RU chassis create a photonic fabric appliance with 33TB of unified memory connected to 16 xPUs through a 115 Tbps electrical network switch. From the GPU's perspective, it has terabytes of high-speed memory even though only a tiny fraction is HBM.
This is a category beyond what CXL alone can deliver. CXL is bounded by copper; once you need to scale memory beyond a single rack, you hit the physical limits of electrical signaling. Optical interconnects (Celestial's Photonic Fabric) are the only way to extend memory pooling to multi-rack scale-up architectures. Marvell did not just acquire technology — they acquired the future of scale-up AI networking. PENG kept its seat at the table as the appliance integrator.
The Honest Timeline Caveat
Marvell's CFO said on the Q3 FY26 earnings call that Celestial AI is expected to add ~$50M annual operating expenses post-closing and become accretive only once meaningful revenue begins in H2 FY2028. The first earnout milestone requires $500M cumulative Celestial revenue by end of Marvell FY29. PENG's photonic memory appliance is on the same multi-year timeline. Bullish color, but anyone telling you 2026 photonic revenue moves the needle is selling. The real story for the next 12 months is CXL/MemoryAI, not photonics.
Top line missed. Bottom line was a knockout.
Q2 FY26 (ended February 2026, reported April 1, 2026) was the most important quarter in Penguin Solutions' recent history because it confirmed the pivot is working at the unit economics level. The headline revenue number looked bad. Everything underneath it looked great.
| Metric | Q2 FY25 | Q2 FY26 | Δ YoY |
|---|---|---|---|
| Net Sales | ~$365M | $343.0M | −6% (vs $374M est) |
| Integrated Memory | ~$105M | ~$172M (50%) | +63% |
| Advanced Computing | $200M | $116M | −42% |
| Non-GAAP Gross Margin | 30.8% | 31.2% | +40 bps |
| Non-GAAP Op Income | ~$33M | $45.3M | +37% |
| Adjusted EBITDA | ~$36M | $50M | +39% |
| GAAP Diluted EPS | $0.09 | $0.58 | +544% (incl. $27M Celestial gain) |
| Non-GAAP Diluted EPS | $0.42 | $0.52 | +24% (vs $0.37 est, +41% beat) |
| Operating Cash Flow | — | $55M | — |
| Cash + ST Investments | — | $489M | Net cash positive |
| Debt | — | $450M | No maturities until 2029 |
FY2026 Guidance — Doubled
| Metric | Prior Guide | New Guide (Apr 1) | Implied |
|---|---|---|---|
| Net Sales Growth | +6% | +12% | ~$1.55B revenue |
| Memory Segment | +10-20% | +65-75% | $700M+ memory revenue |
| Advanced Computing | flat | −15% to −25% | Deliberate exit |
| LED Segment | flat | −5% to +5% | Stable |
| Non-GAAP EPS | $2.00 | $2.15 | +15% YoY |
| GAAP EPS | $0.85 | $1.30 | +53% YoY |
| Full-Year Gross Margin | ~32% | ~28% (CUT) | Memory mix shift |
The revenue miss came entirely from Advanced Computing being lumpy and Penguin Edge winding down — both deliberate. Memory blew through expectations. The pricing environment for DRAM is favorable and demand from telco, networking, and AI-driven customers is strong. CFO Nate Olmstead said on the call that the only constraint to hitting the high end of the memory guide is securing materials, not finding customers.
Non-GAAP EPS of $0.52 beat consensus of $0.37 by 41%. GAAP EPS of $0.58 came in dramatically above the $0.19 GAAP estimate, but a chunk of the GAAP beat came from the $27M gain on the Celestial AI equity stake — a one-time non-operating benefit, not recurring earnings power. Strip that out and the operating beat is still solid but more modest.
Trade ideas like this, before they hit the timeline.
Join Discord →Three competitor categories. One real differentiator.
PENG sits in an awkward spot — it competes with three different categories of company depending on which segment you're looking at. Here's the honest comparison:
| Competitor | Overlap | Strength vs PENG | Where PENG Wins |
|---|---|---|---|
| Super Micro (SMCI) | AI/HPC server assembly, GPU clusters | Massive scale ($5B+/quarter), deep NVIDIA relationship, hyperscaler revenue | Cleaner story (no DOJ overhang), vertical integration into memory layer |
| Dell Technologies (DELL) | Enterprise AI infrastructure, NVIDIA ecosystem | Enterprise sales channel, scale, services revenue | Specialized in inference memory; PENG/Dell/Deepgram is partnership not competition |
| Hewlett Packard (HPE) | HPC clusters, government/defense AI | Cray brand, exascale credibility, sovereign AI relationships | Faster product cycles, more focused on inference memory specifically |
| Micron (MU) | DRAM modules, SSDs at component level | Vertically integrated DRAM fab, massive scale | PENG buys from MU and adds CXL/system value — supplier not competitor |
The Real Differentiator: Memory + System Integration
SMCI builds servers. Dell builds servers. HPE builds servers. None of them have the SMART Modular memory franchise that gives PENG vertical integration into the memory layer of the AI stack. The MemoryAI KV Cache server is not something Supermicro can ship next quarter — it requires deep memory expertise, supplier relationships, and the certified CXL software stack PENG has been building for years.
That said, PENG is the smallest of these competitors by an order of magnitude. SMCI does $5B+ in a single quarter. PENG does $343M. The vertical integration argument matters, but scale matters more if hyperscalers commoditize the inference memory category. The bull case is that PENG occupies a defensible specialist niche; the bear case is that one of the giants decides this category is worth owning and PENG gets crushed in 18 months.
Customer Validation
Recent specific wins from the Q2 FY26 call and disclosures:
| Customer | Deployment | Significance |
|---|---|---|
| Tier 1 Financial Institution | CXL-powered KV Cache servers, on-premise AI factory | Validates inference use case for high-value enterprise |
| Deepgram + Dell | Voice AI infrastructure collaboration | Partnership, not direct sale — scales channel |
| Georgia Tech | AI Makerspace deployment | Education sovereign AI footprint |
| Korean operator (1,000+ Blackwell GPUs) | Single-cluster integration | Sovereign AI / international diversification |
| Sandia National Labs | NextSilicon accelerator deployment (Vanguard program) | Defense / national lab credibility |
| SK Telecom + SK hynix | Trilateral AI Data Center collaboration (CES 2025) | $200M strategic investment by SKT (Dec 2024 close). SK hynix + SMART Modular jointly developing next-gen memory. Asia-Pacific / Japan / Middle East expansion. |
Five new AI/HPC customer logos in Q2 alone, bringing the first-half FY26 total to seven versus three in the prior-year period. The customer base is broadening fast away from hyperscaler concentration, which is exactly what management has been pitching as the strategic shift.
The street is uniformly bullish. Cash hours just confirmed the move.
Sentiment analysis of $PENG mentions across the latest and top X posts from May 5-7, 2026 — covering the after-hours $40 print on May 5, the gap-up open on May 6, and the May 7 fresh 52-week high of $39.93. Sample of ~22 unique high-engagement posts. Zero bearish or skeptical posts in the high-bookmark sample.
Top Engagement Posts This Week
Recurring Bull Themes in High-Engagement Posts
- "AMD EPYC + memory stack" thesis — connection to AMD server demand commentary and the integrated memory franchise. Originated by @quantLR, amplified across the entire bull discussion.
- CXL + photonic memory differentiation — repeated framing as a specialty memory module designer (not raw DRAM), benefiting from supply tightness and AI/HPC infrastructure spend.
- "Monster re-rate" / "just the beginning" — narrative dominates the @ThematicTrader and follow-on posts. Frames current price action as a real pivot recognition, not a meme run.
- Options gamma squeeze — heavy OTM call buying flagged by @penny_ether and others. Low float + options pressure cited as the reason the move can extend violently.
- SK Telecom "penguin emoji" investment hint — fun meme angle around the SK Telecom + SK hynix collaboration. Adds retail engagement layer to the institutional thesis.
- "$40 confirmed, still early" — multi-account framing positions the May 7 cash-hour high as a milestone, not a top. Continuation expected toward $45-50.
What This Tells Me
The bullish consensus is broad, fundamental-backed, and expanding fast. Discussion volume surged sharply alongside the price move — not the other way around. That's the right sequence for a real rerate (price action attracts narrative, narrative recruits new buyers, new buyers extend the move) versus a pump-and-dump (narrative pushes price, then evaporates).
Earlier April mentions were quieter and more speculative. The shift to volume-backed bullish chatter in May coincides with the AMD server commentary, the May 5 after-hours $40 print, and now the May 7 cash-hour confirmation at $39.93. The fact that the gap-up didn't fade — and instead extended into a fresh 52-week high in regular trading — is the textbook signature of a real institutional rotation, not a momentum head-fake. This is a name moving from undiscovered small-cap to mainstream AI infrastructure trade. Engagement metrics suggest more eyes are coming, not leaving.
Real pivot. Real catalyst. Cheap valuation. Free option.
Bull Case
- Memory growing 63% YoY at 50% of revenue. Full-year memory guide raised from 10-20% to 65-75%. Pricing strong, demand strong.
- CEO pivoting decisively. Killing hyperscaler concentration on purpose. Five new non-hyperscale wins in Q2 alone, 7 first-half vs 3 prior year.
- Inference shift is structural. CXL memory is the right product at the right time as AI moves from training to deployment.
- Marvell-backed photonic option. PMA is the 2027-28 catalyst nobody is modeling. Free option attached to current price after Feb 2 close.
- Net cash position. $489M cash, debt $450M, no major maturities until 2029. Buyback active ($32M in Q2).
- Margin trajectory improving (Q2 specific). Non-GAAP gross margin up 40 bps YoY, op margin 13.2%, EBITDA margin 14.6%.
- Bottom-line beat was massive. Q2 EPS $0.52 vs $0.37 estimate (+41% beat). Real operating leverage in the model.
- Diversified customer base now. Tier 1 financial, Georgia Tech, Deepgram + Dell, SK Telecom + SK hynix, Sandia Labs, sovereign AI customers.
- SK Telecom + SK hynix collaboration. Memory supply chain integration in Asia + channel partnership.
- EV/Sales just 1.25x. AI infrastructure peers trade 3-5x. Pure rerate runway as the AI memory story gets recognized.
- Breakout confirmed in cash hours. May 5 after-hours $40+, May 7 fresh 52-week high $39.93. Gap-up did not fade — institutional buying extending the move.
Bear Case
- Q2 revenue MISSED on the headline. $343M vs $374M estimate. Mix was the story but a top-line miss is still a top-line miss.
- Advanced Computing fell 42% YoY. The pivot is intentional but it's still a real revenue hole that has to be backfilled by Memory.
- Memory is commodity-adjacent. 28-31% gross margins are not software margins. DRAM cycles will turn eventually.
- Full-year gross margin guide CUT. Mix is shifting to lower-margin memory + AI hardware. Higher revenue ≠ higher quality revenue.
- Photonic timeline is 2027-28. Marvell expects PMA-related revenue accretion only in H2 FY28. Don't pay for it today.
- Beta is 2.82. Highly volatile small-cap. Thin float means options and momentum can shake out positions fast.
- Trailing financials are weak. 28.6% trailing GM is low for semis. FCF margin ~9%. The forward story has to deliver.
- Scale disadvantage. Larger AI server vendors do $5B+/quarter vs PENG's $343M. If a giant decides CXL inference is worth owning, PENG gets crushed.
- Momentum names mean-revert. A 76% six-month run with a $39.93 print right at the prior all-time-high zone is textbook profit-taking territory. Even confirmed breakouts often retest the breakout level (~$33-35) before continuation.
The breakout is confirmed. Path to $60+ is open.
Honest scenario map. With the May 7 cash-hour print of $39.93 confirming the May 5 after-hours move, "watching, not chasing" is permanently behind us. The market is rerating this in real-time and the prior breakout level is now support. Here's where it can run from $39.72:
Why The Math Supports The Move
FY26 revenue guide is ~$1.55B. At $39.72 share price, market cap ~$1.98B = 1.28x sales. That's still cheap. Get to 2.0x sales (still below most AI infrastructure peers) and you're at $62+ on FY26 numbers alone. Get to 3.0x sales (where pure-play AI memory and photonics names trade) and you're at $93+ on FY27 numbers.
The setup that makes this differ from a normal momentum chase: the multiple expansion math actually works against current revenue, not future hopes. PENG doesn't need to grow into a stretched multiple. Memory growth at 65-75% does the math for you.
The thesis got confirmed. The chart just gave permission — twice.
Penguin Solutions is the cleanest small-cap memory pivot story in the AI infrastructure stack. The company that used to be SMART Global Holdings has a new CEO, a new "AI factory platform" positioning, and a memory business that's printing 63% YoY growth at 50% of revenue. The MemoryAI KV Cache server is the right product for the inference shift. The Marvell-backed Photonic Memory Appliance is a real free option for 2027-28. The balance sheet is strong — $489M cash, debt down to $450M with no major maturities until 2029, dilution risk is zero, and the buyback is active ($32M of repurchases in Q2 alone).
What changed on May 5-7: the after-hours $40 print on May 5 was the first signal. The May 7 cash session — where the stock printed a fresh 52-week high of $39.93 and held above $39 — was the confirmation. The market is rerating this from "small-cap server company" to "AI memory infrastructure pure-play." That rerate has multi-bagger headroom because the starting valuation is so cheap relative to peers. AI infrastructure names trade at 3-5x EV/Sales. PENG sits at ~1.45x even after the move. The pricing has only just begun to catch up to the story. X sentiment confirms the same thing — the bullish consensus expanded sharply alongside the price action.
The honest version: yes, Q2 revenue missed on the headline. Yes, Advanced Computing is still bleeding 42% YoY. Yes, full-year gross margin guide was cut. None of that matters when memory at 50% of revenue is growing 65-75% and the operating profit line is exploding higher. The market is paying for forward operating profit, not trailing gross margin. EPS guide raised from $2.00 to $2.15. GAAP guide from $0.85 to $1.30. Those numbers are what drive the multiple, not the segment that's being deliberately killed off.
This is the cleanest small-cap setup we've worked through. The thesis is real, the catalyst happened in real-time, and the chart broke out of multi-year accumulation with cash-session confirmation. The rerate is starting, not finishing.
Long PENG. Breakout confirmed in cash hours.
The May 5 after-hours $40 print was the trigger. The May 7 fresh 52-week high at $39.93 is the confirmation — institutional positioning has rotated into the name and the gap-up didn't fade. Position sizing reflects conviction: starting position 4-5%, willing to add to 7-8% if it consolidates above $36 and breaks $40 on volume. Trim 25% at $50 (next round number above breakout). Trim another 25% at $65 (base-case landing zone). Hold the runner for $90+ on photonic catalysts in 2027-28. Hard mental stop on any close below $32 with volume (gap-fill plus margin of safety). The Q3 FY26 print (around July 7-14, 2026) is the next confirmation point — if memory growth holds at 60%+, the rerate continues. The setup is clean. The runway is real. Position into the move.