Equity research

Funda AI: "Our interviews suggest that frontier labs remain committed to training and advancing model capabilities internally, while additional evaluation and remediation may slow the pace of public releases ONLY."

One of the employees they interviewed, including the head of sales at one of the labs said:

"A slowdown aimed at improving safety is actually making overall compute demand increase, not decrease."

$Alphabet(GOOGL.US) $Meta Platforms(META.US) $Microsoft(MSFT.US) $Amazon(AMZN.US) $Oracle(ORCL.US) $NVIDIA(NVDA.US) $AMD(AMD.US)

TrendForce: Global Power Demand Capacity

> Global data center power demand capacity is expected to expand from 122.9 GW in 2025 to 161 GW in 2026.

> The share of AI servers is likely to exceed 40% by 2027, whereas the share of general servers is expected to decline from around 40% historically to 25.6% by 2027.

> AI servers accounted for approximately 25% of total data center power demand capacity in 2025, and are projected to rise to 33.4% in 2026.

> The share of AI servers is likely to exceed 40% by 2027, whereas the share of general servers is expected to decline from around 40% historically to 25.6% by 2027.

> The global data center power supply-demand gap in 2030 is estimated at 268 GW.

> Based on current grid data, the U.S. data center power supply-demand gap is estimated to exceed 170 GW by 2030.

$NVIDIA(NVDA.US) $GE Vernova(GEV.US) $AMD(AMD.US) $Alphabet(GOOGL.US) $Amazon(AMZN.US) $Microsoft(MSFT.US) $Dell Tech(DELL.US)

Samsung Securities: Hyperscaler Compute Business Models

Hyperscalers operate through two primary methods: 1) Computing rental and 2) Model inference services, both of which can achieve a Gross Profit Margin (GPM) of 68% to 81%.

> Model Inference Advantages: Model inference services feature a large revenue base and high margins, which is why Meta focuses on deploying its capacity toward in-house services rather than plain rentals.

> Value of Compute Scarcity: As seen in examples like SpaceX, simple rental services based on compute scarcity can achieve revenue and margin levels comparable to model inference services.

> Widening Gap: The flexibility and profitability gap between operators capable of deploying and servicing massive compute infrastructure versus those that cannot is expected to widen further.

Chart Breakdown

> Compute Rental Service GPM (Left Chart): Shows $21.7B in revenue, $6.9B in cost of goods sold (COGS), resulting in $14.8B in gross profit, yielding a 68% GPM.

> Open-Weight Model Inference Service GPM (Middle Chart): Shows $36.6B in revenue, $6.9B in COGS, resulting in $29.7B in gross profit, yielding an 81% GPM.

> Hyperscaler GPM Comparison by Compute Business (Right Chart): Compares compute rental (GPM 68%, $14.8B gross profit), model inference with 35% training (GPM 71%, $16.9B gross profit), and model inference with 0% training (GPM 81%, $29.7B gross profit).

$Alphabet(GOOGL.US) $Amazon(AMZN.US) $Meta Platforms(META.US) $Microsoft(MSFT.US) $SpaceX(SPCX.US) $Oracle(ORCL.US) $Nebius(NBIS.US)

Samsung Securities: Hyperscalers to have a temporary FCF deficit period spanning from 2Q26 to 3Q28, followed by a dramatic expansion in FCF from 2029 through 2030.

> CapEx Growth as a Rational Investment for RPO: The increase in capital expenditure is driven by rapid growth in Remaining Performance Obligations (RPO)—long-term signed contracts—requiring front-loaded CapEx to fulfill future revenue.

> Strong Future Visibility: For instance, Amazon highlighted that a substantial portion of its additional capacity for 2027 is already contracted, with considerable portions of 2028 capacity also pre-booked.

> Break-Even and Payoff Timeline: Server and network investment break-even occurs in under 3 years on average, allowing for meaningful FCF generation for the subsequent 2 to 3 years.

$Meta Platforms(META.US) $Amazon(AMZN.US) $Dell Tech(DELL.US) $NVIDIA(NVDA.US) $Microsoft(MSFT.US) $Oracle(ORCL.US) $Alphabet(GOOGL.US)

Goldman Sachs: companies we visited see pricing competition remaining healthy, with 800G optical module pricing decline within single digits in 2027, and 1.6T optical module price to sustain at $700+.

$Coherent Corp.(COHR.US) $Lumentum(LITE.US) $Applied Optoelectronics(AAOI.US) $Macom Tech(MTSI.US)

Morgan Stanley has CapEx of CSPs ~12% in 2028, with spending on AI semiconductors expected to increase even more.

> CoWoS/CoPoS/EMIB-T capacity is estimated to grow by 50% in 2028, while infrastructure spending (such as factories and AI data centers) will remain flat, and memory prices will grow moderately in 2028.

> TSMC's CoWoS capacity will expand from 130,000 wafers per month in 2026, 200,000 wafers per month in 2027, to 260,000 wafers per month in 2028.

> MI500 series (codename Arcadia) is expected to enter small-scale production by the end of 2027. Production capacity for both the MI455 and MI450 is likely to reach approximately 500,000–650,000 units.

> TSMC's total CoWoS-L orders will increase by 200% year-on-year to 210,000 wafers in 2027.

> TSMC's CoPoS capacity is projected to reach 50,000–100,000 wafers per month in 2028 and 100,000–200,000 wafers per month in 2029, but this will depend on the results of the trial production.

> MS forecasts that TSMC's CoPoS capacity is projected to reach 50,000–100,000 wafers per month in 2028 and 100,000–200,000 wafers per month in 2029, but this will depend on the results of the trial production.

$Taiwan Semiconductor(TSM.US) $AMD(AMD.US) $Intel(INTC.US) $Alphabet(GOOGL.US) $Amazon(AMZN.US) $Meta Platforms(META.US) $Microsoft(MSFT.US)

Evercore ISI: AI networking to be the highest growth segment within AI data center semis the next 5 years.

$Semtech(SMTC.US) $Credo Tech(CRDO.US) $Corning(GLW.US) $Nokia Oyj(NOK.US) $Ciena(CIEN.US) $Lumentum(LITE.US) $Coherent Corp.(COHR.US)

Preliminary Exports | September 10, 2026

Based on figures from the 1st to the 10th of each month

DRAM: +529% Y/Y, +79% M/M

MCP (HBM): +293% Y/Y, +87% M/M

MLCC: +38% Y/Y, +65% M/M

$Micron Tech(MU.US) $DRAM $DISK $EWY $SK Hynix(SKHY.US)

KB Securities sees Samsung boost HBM4 share and foundry yields in South Korea

> Samsung's 33% share of the HBM market in the second quarter this year, is expected to approach a 40% share in the fourth quarter.

> HBM4 sales in the third quarter will increase more than threefold from the previous quarter, and HBM4 is expected to account for more than 60% of total HBM sales in the second half.

> Samsung's foundry has stabilized its 4-nanometer Production yield at over 80%, while the 2-nanometer GAA yield has quickly improved to over 70%, restoring mass-production competitiveness in leading-edge processes.

Jensen says its not circular financing because they only put a little bit of money in to get a lot of money back.

"We put in 1 and 100 comes back in. Is that circular? If that is, let's do more of that."

$NVIDIA(NVDA.US)

Goldman Sachs: expect rack-level AI servers powered by Nvidia at 50k/92k/148k racks in 2026-28E, and AMD at 5k/13k/15k units in 2026-28E.

Almost 10x more racks than AMD by Nvidia in 2028 🔥

$NVIDIA(NVDA.US) $AMD(AMD.US)

MIC report with a supply/demand assessment on various AI components.

HBM/Commodity DRAM: tight through 20227-2028

High-end MLCC: tight in 2027; 2028 to be watched

ABF Substrate: may ease in 2028

ABF Buildup Film: tight in 2027; may ease in 2028

High-end CCL: tight in 2027; may ease in 2028

T-Glass: possible turning point in H2'27

NAND: may ease from late 2027 through 2028

$Micron Tech(MU.US) $DRAM $EWY $DISK $Sandisk(SNDK.US)

Hock Tan with probably the worst explanation ever in countering the circular financing claim:

"Of the six, two of them, two of the strongest guys do not have the cash flows to be able to generate the compute needed to train and create inference, productize their models."

"This is not circular financing. When a customer sees the demand that drives up their own demand and they do not have to use us necessary, they can get it somebody else, that is not circular finance. We are using financing to create demand."

$Broadcom(AVGO.US)

Qualcomm's EVP, CFO, and COO, Akash Palkhiwala, says they have another hyperscaler engagement similar to Amazon:

"we have a very strong engagement with another hyperscaler that we have not disclosed, but similar to the technology engagement we have with Amazon."

$Qualcomm(QCOM.US) $Microsoft(MSFT.US) $Meta Platforms(META.US) $Alphabet(GOOGL.US) $Amazon(AMZN.US)

CVP of Financial Strategy and Investor Relations at AMD, Matt Ramsay, says copper will continue to be used for quite some time:

"I do not think you will see it be a point-in-time thing where everybody just decides as an industry, okay, no more copper, now we are going to go optics."

"You should think about copper and optics living side by side for a number of generations."

$Semtech(SMTC.US) $Credo Tech(CRDO.US) $Amphenol(APH.US) $Lumentum(LITE.US) $Coherent Corp.(COHR.US) $NVIDIA(NVDA.US) $AMD(AMD.US)

According to @FundaAI, GPT-6 used roughly 10x the training compute of the GPT-5 generation.

"Based on our industry discussions, experiments, RL, synthetic data generation and supporting infrastructure can together consume several times as much compute as the main pre-training run, potentially as much as ten times."

"Once TPU 8t and Vera Rubin ship in volume, the next pre-training acceleration can begin, and large-scale interconnect and optics benefit most in that cycle."

$NVIDIA(NVDA.US) $Alphabet(GOOGL.US)

Citi: Scale-out should be one of the major scenarios in

the global datacom optical interconnects market due to

expanding AI datacenter; scale-up should emerge and

become vital

$Lumentum(LITE.US) $Coherent Corp.(COHR.US) $NVIDIA(NVDA.US)

IQE reiterating what the industry has said about InP:

"As we've previously highlighted, we are seeing high demand for our indium phosphide and gallium arsenide optical interconnect technologies."

"Indium phosphide substrates are becoming some of the bottlenecks around significant volume increases."

$AXT(AXTI.US)

Goldman Sachs: Global Optical Transceiver Shipments

> Shipment & Forecast Revisions: Goldman Sachs raised global optical transceiver shipments for 2026–28E by 21%/31%/31%. For 1.6T and above, shipments were raised by 29%/61%/50% for 2026–28E, driven by higher rack-level/ASIC server shipments, increased optical module usage per GPU, and specification upgrades.

> Market Size Projections: The global optical transceiver market is expected to reach US$68bn in 2026E, US$131bn in 2027E, and US$148bn in 2028E. The 800G and above segment is projected to grow at a +69% CAGR to reach US$45bn, US$108bn, and US$130bn respectively.

> Volume and Speed Milestones:

800G/1.6T shipment volumes are projected at 45m/33m units in 2026E, growing to 49m/71m units in 2027E.

3.2T shipments are expected to ramp up to 23m units in 2027E and 68m units in 2028E.

> Technology Adoption: Silicon photonics is expected to account for 60% of 800G, 80% of 1.6T, and 80% of 3.2T optical transceivers.

$Coherent Corp.(COHR.US) $Lumentum(LITE.US) $Applied Optoelectronics(AAOI.US) $Macom Tech(MTSI.US) $NVIDIA(NVDA.US)

MS Hwang from Counterpoint Research believes YMTC could reach 18% global NAND market share in 2028.

> 2028 could be an inflection point in pricing dynamic

> Demand is expected to remain stable through 2028 due to AI-driven growth

> Chinese firms are increasing their tool localization rate

$Sandisk(SNDK.US) $EWY $DRAM $DISK

Several important names reported this week and last week, including Nvidia, Dell, HPE, Broadcom, Ciena, and Credo.

Almost all of them mentioned how demand exceeds available supply, hindering their growth for the following quarters.

Looking through their earnings call, almost all of the compute/server companies mentioned memory and substrate as a significant supply constraint.

Broadcom specifically mentioned "demand for lasers, whether it's EML lasers, CW lasers, is far surpassing supply out there in the industry."

Moving over to optical, Ciena reported LTAs through 2029 for certain optical components. Same signal from Credo.

$Broadcom(AVGO.US) $NVIDIA(NVDA.US) $Hewlett Packard Enterprise(HPE.US) $Ciena(CIEN.US) $Dell Tech(DELL.US) $Credo Tech(CRDO.US)

Morgan Stanley: WFE

A lot of growth coming in memory and foundry/logic WFE in 2026 and 2027.

$Lam Research(LRCX.US) $Onto Innovation(ONTO.US) $ASML(ASML.US) $MKS(MKSI.US) $Sunnova Energy(NOVA.US) $Applied Materials(AMAT.US)

Morgan Stanley: server share of DRAM demand to rise from 37% in 2023 to 59% in 2028e.

$NVIDIA(NVDA.US) $DRAM $Micron Tech(MU.US) $EWY $SK Hynix(SKHY.US) $Dell Tech(DELL.US)

Morgan Stanley: Where are CXMT and YMTC positioned in memory?

The big 3 remain to have strong presence in every memory type.

CXMT and YMTC are still behind by a mile.

$Micron Tech(MU.US) $DRAM $EWY $SK Hynix(SKHY.US) $Sandisk(SNDK.US)