Expert Takeaways: Amazon AWS
> Incremental physical AI capacity stands at 55% AWS Trainium versus 45% NVIDIA GPUs.> Projected net-add AI capacity of 7.3 GW breaks down physically to 60% Trainium and 40% NVIDIA hardware, translating to a roughly 50/50 capital-expenditure split due to NVIDIA's ~1.5x higher per-GW capital cost.> Current Trainium capacity is split 60% inference and 40% training, with inference share expected to rise to 70% in 2027 as legacy Trainium 2 clusters are repurposed and pre-training jobs migrate to Trainium 4 clusters toward late-2027.$Amazon(AMZN.US) $NVIDIA(NVDA.US)
Equity research
E
J.P. Morgan: server compute memory revenue mix comparison across major memory makers
$Sandisk(SNDK.US) $Micron Tech(MU.US) $SK Hynix(SKHY.US) $DRAM $EWYE![图片 1，共 1 张]()
If anyone is curious about the pace of deterioration AI infrastructure names are seeing from their 52 week high. Here is a table of companies in the WFE, accelerators, neoclouds, optical, and memory market.
Best performer? Nvidia, -25% Worst performer? Applied Optoelectronics, -80% This is a non-exhaustive list. $Micron Tech(MU.US) $NVIDIA(NVDA.US) $Lumentum(LITE.US) $AMD(AMD.US) $Dell Tech(DELL.US) $Broadcom(AVGO.US) $Lam Research(LRCX.US)E![图片 1，共 4 张]()
![图片 2，共 4 张]()
![图片 3，共 4 张]()
![图片 4，共 4 张]()
Jeongdong Choe, Senior Vice President of the Semiconductor Division at TechInsights, expects the memory boom to last for 3-4 more years.
> Supply currently meets only about 85% of demand. > DRAM supply is short by approximately 10–12%, with NAND facing a similar deficit.> SVP Choe anticipates that supply levels will rise to 90–95% of demand by next year. Given that a balanced market typically requires supply to exceed demand by 2–3%, a seller's market is likely to persist.> A potential supply-driven downturn might re-emerge in 5–6 years, but that remains a long-term risk not currently weighing on short-term outlooks.> Supply at 100% of demand is still technically tight; true equilibrium requires supply at 102–103%.> In the second half of next year, as HBM4 adoption ramps, Samsung's market share could rise to 25–30%.> In DRAM, the process gap between Korean and US producers has largely closed.$Micron Tech(MU.US) $DRAM $EWY $SK Hynix(SKHY.US) $Sandisk(SNDK.US)E
Nomura: we estimate the 800G/1.6T optical transceiver shipments will grow from 55.0mn/25.1mn units in 2026F to 112.0mn/104.5mn units in 2028F, which should be the key driver for InP substrates.
$Coherent Corp.(COHR.US) $AXT(AXTI.US) $Applied Optoelectronics(AAOI.US) $Lumentum(LITE.US)E![图片 1，共 1 张]()
Rosenblatt: "We believe each $1mn of InP substrates can support ~$30mn of transceiver sales, so the $194.4mn 2027 contract could underpin ~$5.8bn of annual transceiver revenue, or roughly 12mn modules at the ~$500 blended ASP implied by AAOI's 2027 revenue and capacity targets. Over the full five-year agreement, that implies ~$29bn of transceiver revenue at constant pricing."
$Applied Optoelectronics(AAOI.US)E![图片 1，共 1 张]()
Nomura: laser chip shipment volume is forecasted to grow from 480 million units in 2026 to 966 million units in 2027, and reach over 1.5 billion units by 2028. CW laser experiences the largest expansion, growing by 402 million units (from 171 million in 2026F to 573 million in 2028F).
$Coherent Corp.(COHR.US) $Lumentum(LITE.US) $Applied Optoelectronics(AAOI.US) $Broadcom(AVGO.US)E
SemiAnalysis sees 30.3 GW of planned 2028 capacity as risk to moratorium if the candidate proposing a datacenter pause won in every state where one is running.
"If the candidate proposing a datacenter pause won in every state where one is running, those states would together contain roughly 30 GW of planned capacity scheduled for 2028. How much actually slips depends on the permits each project still needs and how long the pause lasts."$Meta Platforms(META.US) $Alphabet(GOOGL.US) $Amazon(AMZN.US) $Microsoft(MSFT.US) $Oracle(ORCL.US) $SpaceX(SPCX.US)E![图片 1，共 1 张]()
LightCounting: Although we expect scale-up switching will grow fastest, Ethernet will still deliver strong sustained growth with a projected 28% CAGR. All told, 2026 data-center switch sales are projected to grow 162% year-over-year in aggregate, and our forecast shows a 5-year CAGR of 36%. By 2031, Ethernet, InfiniBand, and scale-up switch ASICs should generate a combined $147 billion in sales.
$NVDDA $Broadcom(AVGO.US) $Alphabet(GOOGL.US)E
Citi: Memory
"Overall, we believe the market is underestimating how constrained DRAM/HBM supply is likely to remain as AI-driven demand from hyperscalers continues to accelerate, while memory makers face diminishing benefits from process shrinks and must rely on long-cycle greenfield capacity additions. As a result, HBM and advanced memory supply could remain tight through 2027 and 2028.""Supply is the underappreciated piece, in our view: Node transitions yield less density (1-delta ~25% vs 1-gamma and prior nodes 30-40%). HBM's trade ratio eats wafer capacity, and DRAM bit growth is only in the low 20s. Real supply relief needs greenfield fabs, which takes a long time and cleanroom space is the constraint. Micron's construction spend as a % of capex is approaching 50%."$Micron Tech(MU.US) $DRAM $EWYE
Citi: According to LightCounting and Frost & Sullivan, the 800G segment achieved an exceptional CAGR of 156.3% from 2021 to 2025 and is forecast to grow at 13.9% from 2025 to 2030E. The 1.6T segment — representing the next generation — is forecast to grow at a CAGR of 98.9% from 2025 to 2030E, with mass adoption forecast to begin in 2026.
$Lumentum(LITE.US) $Coherent Corp.(COHR.US) $Applied Optoelectronics(AAOI.US) $Macom Tech(MTSI.US) $Fabrinet(FN.US)E
Nomura: Based on our latest tracking, incorporating project delays, cancellations and additions, we now have 51 GW-scale data center projects in our tracking base.
Our latest calculation indicates 26GW/40GW/28GW incremental deployment in 2026F/2027F/2028F, from 27/32/23 GW in June update.We estimate 26GW/40GW/28GW could translate into the demand for 5-8mn AI chips.$NVIDIA(NVDA.US) $AMD(AMD.US) $Alphabet(GOOGL.US) $Meta Platforms(META.US) $Microsoft(MSFT.US) $SpaceX(SPCX.US) $Amazon(AMZN.US)E![图片 1，共 1 张]()
According to @G2AIStack, Nvidia Kyber is back on the agenda, but a scale-down version with the proposed GPU count per rack reduced from 144 to 72.
Additionally, Google plans to ship ~9 million TPU v8 chips in 2027. $NVIDIA(NVDA.US) $Alphabet(GOOGL.US)E
$Sandisk(SNDK.US): "We are in active conversations with existing NBM customers that are evaluating extending existing agreements and with new customers that are evaluating new NBMs."
E![图片 1，共 3 张]()
![图片 2，共 3 张]()
![图片 3，共 3 张]()
Presentation from Celestica highlighting projections for custom ASIC and ethernet networking market.
> Custom ASIC TAM is projected to grown more than 7x from $61 billion in 2026 to $434 billion by 2030 at a 64% CAGR.> Data Center Ethernet Networking TAM is projected to grow from $44 billion in 2026 to $173 billion by 2030, driven by rapid expansion in Scale-Up Back-End networking from $1 billion to $55 billion.$Celestica(CLS.US) $Alphabet(GOOGL.US) $Meta Platforms(META.US) $Microsoft(MSFT.US) $Amazon(AMZN.US) $Lumentum(LITE.US) $Coherent Corp.(COHR.US)E
Goldman Sachs on HBM pricing:
"TrendForce expects blended HBM ASP to increase 122% YoY to $3.7/GB in 2027, driving HBM industry revenue growth of more than 230% YoY to exceed $200 billion. TrendForce said that pricing may vary materially by supplier and customer, with a wide variance where the high-end of the pricing for HBM4 could reach over $5/GB while the low-end closer to $3/GB for next year.""TrendForce believes 8-hi HBM4 could command 5–10% premium over 12-hi on a per-bit basis due to higher base-die cost per memory unit, while pricing for HBM4E next year could be materially higher than HBM4 pricing. We believe TrendForce's view to be consistent with our expectations, as we currently estimate HBM ASP for next year to rise 100-160% YoY for the major suppliers."$Micron Tech(MU.US) $DRAM $EWYE![图片 1，共 3 张]()
![图片 2，共 3 张]()
![图片 3，共 3 张]()
Expert Takeaways: Google
Gemini 4 Argon Launch Focus: Finalized its checkpoint about 1.5 months before release, targeting Claude 3.5 Sonnet on price/performance with legal and financial as key verticals.Severe Compute Constraints at Google: Due to heavy compute shortages, Gemini 4 Argon is globally suspended internally, free users are pushed from Flash to Light, and features like AI Mode Overview are curtailed.Hardware Bottlenecks & Model Demand: TPU v8i delivery is delayed by roughly one quarter, and since Gemini 4 Argon is Google's first unquantized/unreduced large release, it is extremely compute-heavy.Anthropic's TPU Share & Revenue: Anthropic holds about 3.5 million of Google's 7 million+ TPU units, and its GCP rented compute revenue outpaces running Gemini 4 Argon, prioritizing external supply.Aggressive B-End Integration for Anthropic: Google is integrating Anthropic as a "OneP" (First Party) product inside Gemini Enterprise and Workspace with top-level menu placement equal to Gemini.Flexible Token Credits System: C-end and B-end bundles use equivalent token credits (e.g., 1M Anthropic tokens = 2 coins, 1M Argon tokens = 1 coin) with a shared, freely allocatable quota.Anthropic Training & Infrastructure Mix: Anthropic's mix is roughly 40% training and 60% inference, with the Fable series trained on TPU V7 and about half its product line running on TPUs.Antigravity vs. OpenCode Performance: Antigravity suffers from bugs, privacy constraints preventing source code viewing, and lagging feature velocity; Argon plus OpenCode scores higher on Elo and benchmarks.C-End AI Economics & Growth: C-end AI remains loss-making, with Google banning 10k–20k free-riding accounts daily; new model releases every ~2 months drive 10–20% upgrades and double token consumption.Current Infrastructure Bottlenecks: The tightest bottlenecks are power generation equipment for data centers and inference compute (especially memory/cache for coding workloads), while training compute is less scarce due to paused projects.$Alphabet(GOOGL.US)E
Mizuho: We believe hyperscalers such as Google and Meta are exploring CPO architectures and could help drive the roadmap, though our call noted volume ramps only in 2028-30E, with current supply chain bottlenecks.
$Alphabet(GOOGL.US) $Meta Platforms(META.US) $Tower Semicon(TSEM.US) $Lumentum(LITE.US) $Coherent Corp.(COHR.US)E![图片 1，共 1 张]()
Deutsche Bank sees revenue upside of $216.1 billion. for SpaceX exiting 2027E. This comes with an estimated $148.4 billion in future deals.
$SpaceX(SPCX.US)E
Evercore ISI: AEC Demand
> Copper and AECs are expected to stay in-rack through 200G per lane (1.6T applications), while CPO is viewed as a 2028 and beyond story.> Industry experts estimate an attack-rate of 1.5 AECs per GPU/ASIC.> Passive copper cables cost roughly $100, AECs cost about $900, and AOCs range from $3,600 to $4,000.> 400G AECs are becoming more commodity-like, 800G is in mass production, and 1.6T is expected to carry premium pricing.> Credo's core advantage lies in its system approach (owning SerDes/DSP silicon, software, and end-to-end cables/transceivers) and strong incumbency in the AEC franchise.$Credo Tech(CRDO.US) $Maxlinear(MXL.US)E![图片 1，共 1 张]()
Expert Takeaways: Microsoft
> Azure projected YoY growth of ~45% (±200 bps), with quarterly revenue guidance of $34‑35 billion.> Current‑quarter capex is expected ≥$50 billion; 70‑80% allocated to short‑lived assets such as GPUs and accelerators rather than real‑estate‑heavy datacenter build‑outs.> Over 70% of available compute capacity is prioritized for Microsoft’s internal M365 Copilot ecosystem, with residual capacity allocated to external Azure Foundry customers.$Microsoft(MSFT.US)E
BofA 2H26 survey on AI with ~300 CIOs:
"~17% already see AI ROI, additional 31% expect to see it in 12 months As AI projects move into production, the next test is whether they deliver measurable returns. Enterprise-wide ROI is already reported by 17.2% of respondents, with another 31.4% expecting it within 12 months. Combined, 48.6% see current or near-term returns, while 30.7% expect them in one to two years.""91.2% of respondents increased their AI compute use over the past 12 months, and 54.0% increased it significantly."$NVIDIA(NVDA.US) $Alphabet(GOOGL.US) $Meta Platforms(META.US) $Microsoft(MSFT.US) $Amazon(AMZN.US)E![图片 1，共 1 张]()
Expert Takeaways: Memory
> S/D gap of at least 20% in second half of 2026. > Big vendors have just begun forecasting since demand for 2027 has not settled yet.> S/D gap for 2027 is possibly around 10%. A pessimistic take believing in balance by end of next year.$Micron Tech(MU.US) $DRAM $EWY $SK Hynix(SKHY.US)