Every quarter I participate in a call for buy-side investment analysts focused on the broader datacenter/hyperscaler semiconductor ecosystem

Here, I’m sharing the raw notes I used to drive the Q1 2026 call (it’s my cheat sheet).

(If you’d like more information about how to participate, feel free to email us at [email protected].)

Q1 Key Announcements (all vendors)

Company	Deal	Party	Value / Scale	Details
AMD	AI chip supply agreement	Meta	~$60B–$100B potential	Multi-year deal to deploy 6 GW of AMD Instinct GPUs across Meta AI infrastructure. Includes a custom MI450 GPU and Helios rack architecture with EPYC CPUs.
AMD	Strategic chip supply + equity option	Meta	Up to 10% AMD equity via warrants	Meta can receive up to 160M AMD shares tied to GPU shipment milestones.
AMD	AI chip supply agreement	OpenAI	Multi-year (value undisclosed)	OpenAI agreed to purchase 6 GW of AMD AI processors over several years, with a similar warrant structure.
Broadcom	Custom AI accelerator design program	Anthropic	Multi-gigawatt deployments	Broadcom expects to supply 1 GW of custom chips in 2026 scaling to 3 GW by 2027.
Broadcom	Custom AI chip design partnerships	Alphabet, Microsoft, Amazon, Meta	Multi-year programs	Broadcom designs custom AI processors for multiple hyperscalers.
Broadcom	AI chip development partnership	OpenAI	Production expected ~2027	OpenAI is working with Broadcom on its first in-house AI accelerator.
NVIDIA	Strategic ecosystem investment	Lumentum & Coherent	~$4B total investment	NVIDIA investing $2B each in optical networking companies to accelerate photonic interconnects for AI chips.
NVIDIA	Large AI chip supply deal	OpenAI	Part of broader ~$100B AI investment ecosystem	NVIDIA expected to supply AI processors and potentially invest in OpenAI infrastructure.
Marvell	Custom ASIC and networking supply	Amazon, Microsoft (reported by analysts)	Multi-year infrastructure demand	Marvell supplies custom chips and networking silicon used with Trainium and Maia AI systems.
Qualcomm	Data-center AI accelerator push	Cloud providers / AI infrastructure vendors	Product launch cycle 2026	Qualcomm introduced AI200 / AI250 inference accelerators targeting data-center AI workloads.
MediaTek	Edge-AI and telecom AI partnerships	Telecom operators and device makers	Strategic expansion	MediaTek expanding AI chips across 6G infrastructure, edge AI, and automotive platforms.

---

Semi Company Detail

AMD

Category	Event / Announcement	Product / Technology	Key Customers / Partners	Notes
Major hyperscaler deal	Multi-year 6-gigawatt AI infrastructure agreement	Instinct GPUs + EPYC CPUs	Meta	AMD will supply up to 6 GW of Instinct GPUs and EPYC CPUs for Meta’s AI data centers, with first deployments starting in 2H 2026.
Financial structure of Meta deal	Partnership includes equity warrants	Instinct GPU roadmap	Meta	Meta receives warrants that could convert to ~10% of AMD, aligning incentives with AMD’s AI growth.
Earlier hyperscaler deal (still impacting 2026)	Multi-year AI chip supply agreement	Instinct GPUs	OpenAI	OpenAI signed a similar 6 GW supply agreement, positioning AMD as a second supplier to major AI labs.
Rack-scale AI platform launch	AMD unveiled Helios rack-scale architecture	Helios AI rack	Meta (OCP)	AMD’s first rack-scale system competes directly with NVIDIA’s NVL systems and hyperscaler racks.
Next-gen GPU roadmap	AMD previewed MI400-series accelerators	Instinct MI430X / MI440X / MI455X	Cloud providers & hyperscalers	The MI400 family targets next-generation AI training and inference infrastructure.
AI supercluster deployments	Large AI clusters planned using AMD GPUs	Instinct MI450 + Helios racks	Oracle Cloud	Oracle is building AI superclusters with ~50,000 Instinct GPUs.
Next-generation CPU roadmap	EPYC “Venice” server CPUs announced	Zen-based EPYC processors	Hyperscalers and AI data centers	Venice CPUs will pair with MI450 GPUs in AI racks.

Broadcom

Category	Event / Announcement	Product / Technology	Key Customers / Partners	Notes
Explosive AI revenue growth	AI semiconductor revenue more than doubled year-over-year to $8.4B in Q1 FY2026	Custom AI ASICs + networking silicon	Hyperscalers and AI labs	AI demand is now the primary driver of Broadcom’s semiconductor business, fueling 29% revenue growth to $19.3B.
Massive AI chip forecast	Broadcom projected over $100B in AI chip revenue by 2027	Custom XPUs	Hyperscalers including Google, Microsoft, Amazon, Meta
Hyperscaler ASIC programs expanding	Broadcom confirmed six major AI chip customers	Custom AI accelerators	Google, Meta, OpenAI, Anthropic, others
Anthropic deployment	Plans to supply 1 GW of AI chips in 2026, scaling to 3 GW by 2027	Custom AI ASICs	Anthropic
Advanced chip-stacking technology	Broadcom developing 3D-stacked AI chips, targeting 1M units shipped by 2027	Advanced packaging / chip stacking	Fujitsu and other AI customers	Improves bandwidth and power efficiency for AI workloads, enabling larger accelerators.
Networking leadership for AI clusters	Continued growth of Tomahawk Ethernet switches and SerDes technologies	AI data-center networking	Hyperscalers
Next-gen switching roadmap	Development of 1.6T–3.2T networking and terabit switching	DC switching silicon	Cloud and hyperscale infrastructure providers	Next-generation AI clusters will require major upgrades to data-center networking, strengthening Broadcom’s position.
VMware integration progress	VMware Cloud Foundation becoming a major enterprise software platform	Infrastructure software	Enterprises and cloud providers	VMware revenue and subscriptions continue to grow as Broadcom integrates the platform into its enterprise strategy.
Capital allocation	Announced $10B share buyback program alongside strong earnings	Corporate finance	Investors	Confidence in long-term AI demand and cash flow.

Intel

Category	News Item	Details	Notes
AI accelerator roadmap	Falcon Shores canceled	Intel decided not to commercialize the Falcon Shores AI GPU and will use it as a development platform for future architectures.	A reset of Intel’s AI accelerator strategy after delays and competitive pressure.
New AI GPU	Crescent Island inference GPU	Intel unveiled a new AI inference-focused GPU based on Xe architecture with large memory capacity for enterprise servers.	Intel is pivoting toward cost-efficient inference workloads, a massive market expected to outgrow training.
AI partnerships	Technical collaboration with SambaNova	Intel announced a multi-year partnership with SambaNova Systems to advance AI infrastructure.	Is Intel leaning into AI startups to strengthen its ecosystem?
Client AI chips	Core Ultra Series 3 launch	Intel introduced processors for AI PCs built on its 18A manufacturing node.	Shows Intel’s push to lead the AI PC market, where it has stronger OEM relationships.
Next-gen AI architecture	Jaguar Shores roadmap	Intel is developing a future AI platform called Jaguar Shores, expected to follow Gaudi accelerators.	Intel’s longer-term attempt to build rack-scale AI systems.
AI accelerator deployment	Gaudi ecosystem expansion	Gaudi accelerators are available in systems from vendors like Dell, HPE, Lenovo, and Supermicro.	Intel’s main foothold in data-center AI currently remains Gaudi accelerators.
Data-center CPU roadmap	Clearwater Forest Xeon delay	Intel pushed back its next-generation Xeon server processor to 1H 2026 due to packaging challenges.	Ongoing manufacturing and packaging challenges

Marvell

Category	Event / Announcement	Product / Technology	Key Customers / Partners	Notes
Major acquisition (AI connectivity)	Completed $3.25B acquisition of Celestial AI	Photonic Fabric optical interconnect	AI cloud providers, hyperscalers	Expands Marvell’s ability to build optical interconnects for large AI clusters, improving bandwidth and power efficiency in data-center fabrics.
Optical interconnect strategy	Celestial AI technology integrated into Marvell’s data-center portfolio	Silicon photonics + chip-to-chip optical links	Hyperscale AI data centers	Photonic Fabric enables high-bandwidth, low-latency connectivity between AI accelerators
Networking acquisition	Announced $540M acquisition of XConn Technologies	CXL switching and AI chip interconnect	AI infrastructure vendors	XConn’s technology enables AI accelerator pooling and memory disaggregation using CXL fabrics.
Financial outlook tied to AI	Custom AI chips expected to drive 20%+ growth in ASIC revenue	AI ASICs	AWS and hyperscalers	Demand for Trainium-related silicon is expected to drive strong revenue growth in 2026.

NVIDIA

Category	Event / Announcement	Product / Technology	Key Customers / Partners	Notes
Next-gen AI platform launch	NVIDIA unveiled the Rubin AI platform, successor to Blackwell	Rubin GPU + Vera CPU + NVLink 6 fabric	Hyperscalers, AI labs	Rubin is a six-chip architecture designed for AI factories, integrating CPU, GPU, networking, and memory into a single rack-scale system.
Rack-scale AI architecture	Introduction of NVL72 rack systems built on Rubin superchips	Vera Rubin NVL72	Hyperscalers and AI infrastructure providers
Software ecosystem expansion	Expanded collaboration to optimize enterprise AI stack	CUDA + AI software stack	Red Hat	Joint effort integrates Rubin with Red Hat Enterprise Linux, OpenShift, and hybrid-cloud AI tooling.
AI infrastructure investment	NVIDIA committed $4B to photonics companies	Optical networking for AI clusters	Lumentum, Coherent	Investments to accelerate optical interconnects critical for scaling AI clusters and reducing networking bottlenecks.
Strategic optics partnerships	Multiyear agreements for advanced optical networking technology	Silicon photonics, optical transceivers	Lumentum, Coherent	Agreements include multibillion-dollar purchase commitments and capacity guarantees
AI-native telecom ecosystem	NVIDIA joined telecom leaders to build AI-native 6G infrastructure	AI-RAN platforms	Nokia, Cisco, Deutsche Telekom, SK Telecom, others	Expands NVIDIA GPUs into telecom networks as AI-accelerated radio infrastructure.
Networking architecture expansion	NVLink ecosystem expanding with CPU partners	NVLink Fusion interconnect	Intel, Fujitsu, Arm ecosystem	Allows CPU vendors to integrate NVLink, strengthening NVIDIA’s GPU-centric compute ecosystem.
Competitive pressure from custom silicon	Hyperscalers expanding in-house AI chips	Custom ASICs	Amazon, Google, Microsoft, Meta	Growing competition from custom silicon is pushing NVIDIA to strengthen system-level differentiation.

---

Hyperscaler ASIC Business

Company	AI Chip	Primary Use	Key Silicon Partner(s)
Google	TPU	Training + inference	Broadcom + Marvell + TSMC
Meta	MTIA (	Recommenders + generative AI	Broadcom + TSMC
Microsoft	Maia	Azure AI infrastructure	Marvell + TSMC
Amazon (AWS)	Trainium + Inferentia	Training (Trainium) + inference (Inferentia)	Annapurna Labs + TSMC
ByteDance	Custom XPU (unnamed publicly)	Training + inference	Broadcom (rumored)
OpenAI	Custom accelerator (Titan / internal project)	Frontier model training + inference	Broadcom

Google TPU

TPU v7 “Ironwood”

• Ironwood is Google’s seventh-generation TPU—most performant and scalable custom AI accelerator to date, first designed specifically for inference

Production Volume & Manufacturing

• Fubon projects total TPU v7 racks will reach about 36,000 units in 2026
• Total TPU production estimated at 3.1 to 3.2 million units in 2026, constrained mainly by advanced packaging capacity
• In January 2026, Google confirmed that its custom-designed TPUs have outshipped general-purpose GPUs in volume for the first time

Partnership Structure Evolution

• Google is teaming with MediaTek and Broadcom for Ironwood—spreading cost and production risk
• MediaTek handles I/O and manufacturing coordination with TSMC; Broadcom continues developing high-performance cores
• Google selected MediaTek partially because the company offers costs approximately 20–30% lower than alternative partners
• MediaTek focuses on cost-sensitive variants such as the TPUv7e, with production scheduled to begin in 2026

Selling TPUs

• Anthropic has access to as many as 1 million Google TPUs, expected to bring over a gigawatt of new AI compute capacity online in 2026
• Anthropic has committed $21 billion to custom chip orders via Broadcom, securing nearly 1 million Google TPU v7p units for delivery by late 2026
• Meta signed a multi-billion-dollar TPU lease agreement to access chips via Google Cloud for training AI models
• Meta is separately in talks to purchase TPUs outright for deployment in its own data centers starting in 2027

TCO Advantage vs. NVIDIA

• From Google’s perspective, all-in TCO per Ironwood chip for full 3D Torus configuration is ~44% lower than the TCO of a GB200 server
• Even when offering prices to external customers, TPU v7’s TCO is about 30% lower than NVIDIA’s GB200 and about 41% lower than the upcoming GB300

OpenAI “Titan”

Partnership Details

• OpenAI and Broadcom collaboration covers 10 gigawatts of custom AI accelerators
• OpenAI designs the accelerators and systems; Broadcom develops and deploys
• Deployments targeted to start in 2H 2026, completing by end of 2029
• Systems will use Ethernet-based networking

Chip Specifications

• OpenAI’s chip codenamed “Titan” will use TSMC’s N3 process, mass production expected end of 2026
• Second-generation “Titan 2” reportedly planned on more advanced A16 process

Revenue Timeline for Broadcom

• Broadcom CEO Hock Tan told investors he doesn’t expect much revenue from the OpenAI partnership in 2026
• “I appreciate the fact that it is a multiyear journey that will run through 2029”
• Internal push to roll the chip out in Q2 2026 has already slipped to Q3 at the earliest

Meta MTIA

Current Status

• Meta accelerated deployment of MTIA v3, codenamed “Iris”, 3^rd generation custom silicon
• As of February 2026, Iris chips have moved into broad deployment across Meta’s massive data center fleet
• Flagship Iris was designed primarily with assistance from Broadcom
• Fabricated on TSMC’s cutting-edge 3nm process, integrates eight HBM3E 12-high memory stacks with bandwidth exceeding 3.5 TB/s

Setbacks in Training Chip Program

• Meta scrapped an advanced AI training chip after design roadblocks
• The MTIA program has had a history of setbacks, including scrapping an earlier inference chip after it underperformed
• Meta announced a multiyear agreement with AMD on February 24 worth more than $100 billion for MI450 GPUs

2026 Roadmap

• MTIA-2 already in production, slated to debut in H1 2026, built on TSMC’s 3nm with CoWoS-S packaging
• MTIA-3 set for H2 2026 debut, with GUC handling back-end packaging
• MTIA v4 “Santa Barbara” readying for latter half of 2026 with HBM4 memory and liquid-cooling systems exceeding 180kW per rack
• Arke inference-only chip variant developed in collaboration with Marvell (not Broadcom)—representing supply chain diversification

Partner Diversification Risk

• Meta co-developed its MTIA chips with Broadcom, but is reportedly diversifying with Marvell for certain variants
• Meta has committed up to $135 billion in capital expenditures for 2026 to build out AI infrastructure