Summary: Inference is becoming cheap, efficient, and abundant; training is becoming scarce, capital-intensive, and strategically decisive.

Qualcomm’s acquisition of Ventana (1), Meta’s acquisition of Rivos (2), Intel’s acquisition of SambaNova (3), Marvell’s acquisition of Celestial AI (4), and Nvidia’s acquisition of Groq’s inference assets (5) collectively signal a decisive restructuring of data center AI infrastructure.

These moves reflect two shifts:

Inference is becoming a commodity business driven by TCO (not CUDA), favoring RISC-V CPUs, NPUs, and LPDDR/SRAM-centric designs (illustrated by cases 1, 2, 3, and 5).

This transition enables alternative silicon vendors and in-house hyperscaler designs—by Meta, Google, Amazon, Qualcomm, Groq, and SambaNova—to increasingly displace existing GPU-centric inference stacks.

Implication for MTK:

Evaluate the inference server business opportunity—including potential customers, SAM, and impact on ASIC businesses—and assess server technologies across NPU (we have competitive IP but need to scale up), CPU, system architecture, and software.

Training is becoming the next strategic battleground for ASIC vendors, where scalability define leadership. Photonic fabrics is a critical enabler (illustrated by case 4 and major strategic investments by big-techs).

Consequently, vendors such as Marvell, Broadcom, and Alchip are investing aggressively in optical switching, co-packaged optics, and electro-optical integration to remain relevant as training systems evolve beyond electrical interconnect limits.

Implication for MTK:

Invest in photonic technologies—organically, through partnerships, or via M&A—to remain competitive in the data center ASIC market.

Qualcomm buying Ventana, Meta acquiring Rivos, Intel acquiring SambaNova, and Marvell buying Celestial AI signal a decisive shift in the semiconductor industry. These moves are not about expanding short-term revenues; they represent a restructuring of AI infrastructure to address “silicon sovereignty” from Nvidia’s GPU, CUDA, and NVLink ecosystem.

Overview: The Reshaping of AI Data Center Infrastructure

The semiconductor and systems architecture landscape is driven by the explosive growth of AI inference workloads and intensifying competition in foundation-model training. Four forces are converging.

1. AI Inference at Data Center Is Exploding

Google reported processing 1.3 quadrillion tokens per month (≈43 trillion tokens per day), growing exponentially across search, coding, multimodal agents, and enterprise workloads.

Inference is no longer a downstream by-product of training—it has become the dominant consumer of data-center compute capacity.
This shift forces hyperscalers to re-architect inference infrastructure, moving away from general-purpose GPUs toward specialized, cost-optimized silicon.

2. Inference Is Commoditizing — TCO Replaces CUDA to Drive Design

At hyperscale, inference economics dominate architectural decisions. The traditional CUDA + GPU + HBM + NVLink moat is weakening due to high BOM cost, power inefficiency, and vendor lock-in.

The gap is narrowing: architectures based on RISC-V + NPUs + SRAM/LPDDR now deliver sufficient throughput with significantly lower TCO.
Key cost-reduction vectors:
Replacing general-purpose GPUs with domain-specific NPUs to lower area, latency, power, and cooling cost.
Replacing HBM with LPDDR, GDDR, or on-chip SRAM to reduce memory cost and supply issues.
The result is a shift toward tightly integrated CPU/NPU/memory inference silicon, favoring power efficiency over peak FLOPS.

3. Foundation Model Competition Is Converging Toward a Winner-Take-Most Outcome — Driving Extreme-Scale Training

The foundation-model landscape is consolidating around a small number of global leaders, including Google, OpenAI, Anthropic, and xAI.

While the number of top-tier foundation-model developers is shrinking, model scale, training cost, and system complexity are expanding rapidly.
Architectural differentiation at the model level is narrowing, compressing performance gaps and intensifying competition.
As a result, access to massive, reliable, and scalable training infrastructure becomes a decisive competitive advantage, not merely a cost factor.

4. AI ASIC Vendors Are Moving Upmarket Into Training — Where Photonics Becomes Strategic

As inference profit margins compress, AI ASIC vendors are moving upstream into training, where profit margins remain attractive but performance scaling is challenging.

Training workloads require:
Massive scale-up and scale-out compute
Ultra-high-bandwidth, low-latency interconnects
Photonic fabric is emerging as an enabler:
Celestial AI (acquired by Marvell), Lightmatter (backed by Google), and Ayar Labs (backed by Nvidia, AMD, Intel) are advancing optical interconnects that outperform copper in bandwidth-per-watt and latency.
These technologies enable super-scale compute, memory pooling, and next-generation AI architectures.

Slides

Here’s a set of Gamma-ready slide templates based on your outline. I’ve structured them as concise, self-contained markdown blocks optimized for Gamma’s AI slide generator—each includes a title, key message, visuals cues (e.g., tables, bullets, charts), and layout notes for easy import and auto-formatting. Copy-paste directly into Gamma for instant slides.

Slide 1: Acquisitions Restructure AI Infrastructure

Key Message: The semiconductor industry is undergoing a restructuring driven by explosive growth of AI inference and intensified competition of foundation model training.

Company	Target	Status	Rationale	Opportunity	Technology	Value
Qualcomm	Ventana	Announced Dec 2025	RISC-V hedge to Arm; dual-ISA roadmap	Datacenter RISC-V CPUs	Vector/matrix chiplets	Undisclosed
Meta	Rivos	Rumored	Escape Nvidia dependency; in-house silicon	Custom inference accelerators	RISC-V + NPU	~$2B (est.)
Intel	SambaNova	Pending	Boost inference silicon portfolio	Enterprise AI inference	Reconfigurable Dataflow	~$1.6B (rumored)
Marvell	Celestial AI	Announced Dec 2025	Own photonic IP for scale-out AI	Optical interconnect for AI clusters	Co-packaged optics	$3.25B upfront
Nvidia	Groq (IP)	Announced Dec 2025	License low-latency inference IP	Efficient inference for real-time AI	LPU + SRAM stack	Undisclosed

Visual: Highlight “Announced Dec 2025” rows in green; timeline graphic on right.

Slide 2: AI Compute Surge — Two Distinct Waves

Key Message: Training and inference are distinct workloads imposing different compute, memory, interconnect, and driving different system architectures.

Wave 1: Foundation Model Training

Led by: OpenAI, Google, Meta
Requires: 100K+ XPUs, interconnected via scale-up and scale-out
Characteristics: One long-running (~3 months) workload

Wave 2: Foundation Model Inference

Driven by Applications: Searching, Coding, Agents, Content Generation
Requires: Clusters of 100’s XPUs
Characteristics: Billions of concurrent short-lived (~10 mins) sessions

Visual: Side-by-side wave diagrams (e.g., towering training wave vs. explosive inference bursts); bold key diffs.

Slide 3: AI Inference at Data Center Scale Is Exploding

Key Message: Inference is now the dominant consumer of data center compute and is growing exponentially across applications.

Google reported 1.3 quadrillion tokens per month (~43 trillion per day), growing rapidly across search, code generation, agents, and enterprise workloads.
Inference is no longer a downstream by-product of training—it has become the primary compute workload.
Hyperscalers are re-architecting inference infrastructure away from general-purpose GPUs toward specialized, cost-optimized silicon.

Visual: Exponential growth chart (tokens/month); pie chart showing inference > training; bold stats enlarged.

Slide 4: Inference Is Commoditizing — TCO Replaces CUDA

Key Message: Inference is becoming a TCO game, not a TOPS contest.

Old Stack (Nvidia Moat):

CUDA + GPU + HBM + NVLink
High BOM, power inefficiency, vendor lock-in

New Stack:

RISC-V (Meta/Rivos, Qcom/Ventana, Tenstorrent) + NPU + DDR/SRAM + PCIe/UA/UE
Lower power, latency, and TCO

Design Shifts:

Domain-specific NPUs replace general-purpose GPUs
LPDDR (Qcom), GDDR (Tenstorrent), and on-chip SRAM (Groq) replace HBM
Tightly integrated CPU/NPU/memory inference silicon prioritizing power efficiency over peak TOPS.

Visual: Before/after stack comparison icons; TCO savings bar chart.

Slide 5: Foundation Model Competition Drives Super-Scale Training

Key Message: As the number of model leaders shrinks, the ability to scale training infrastructure becomes the core competitive edge.

The foundation model landscape is consolidating around a few global leaders (Google, OpenAI, Anthropic, xAI).
While developer count shrinks, training cost, model size, and system complexity are increasing.
Access to massive, reliable, scalable training infrastructure becomes a decisive advantage—not cost.

Visual: Funnel graphic (many devs → few leaders); upward arrows for cost/size/complexity; leader logos.

Slide 6: AI ASIC Vendors Shift Upmarket

Key Message: As inference profit margins shrink, ASIC vendors are moving into training infrastructure where photonics become strategic.

Training requires: Massive compute scale-up, scale-out, and scale-across; ultra-high-bandwidth, low-latency interconnects.
Photonic fabric is emerging as key enabler:
- Celestial AI (acquired by Marvell)
- Lightmatter (backed by Google)
- Ayar Labs (backed by Nvidia, AMD, Intel)
Photonics unlock super-scale compute and memory pooling for next-gen AI system architectures.

Visual: Photonics network diagram; vendor logos with backing highlights; scale-up pyramid.

Slide 7: Strategic Implications for MTK

Key Message: MTK must act on two fronts: build photonic capability to compete in training infrastructure and capture inference opportunities.

Invest in Photonic Technologies to move upmarket into training and position as a relevant player for existing business.
Evaluate Inference Server Opportunities
- Potential customers, SAM, and impact on ASIC business
- Assess and invest in server-class technologies: NPU (based on mobile NPU and scale-up required), CPU, system architecture, software.

*Visual: Two-pillar graphic (Photonics

Inference); action arrows; MTK logo centered.*

These templates are plug-and-play for Gamma—import as markdown, and it will auto-generate visuals, animations, and layouts. Would you like me to add more visuals specs, adjust for a specific theme, or expand any slide?

Appendix

Slide 1: Acquisitions Restructure AI Infrastructure

Key Message:

Key Message: The semiconductor industry is undergoing a restructuring driven by explosive growth of AI inference and intensified competition of foundation model training.

Company	Target	Status	Rationale	Opportunity	Technology	Value
Qualcomm	Ventana	Announced Dec 2025	RISC-V hedge to Arm; dual-ISA roadmap	Datacenter RISC-V CPUs	Vector/matrix chiplets	Undisclosed
Meta	Rivos	Rumored	Escape Nvidia dependency; in-house silicon	Custom inference accelerators	RISC-V + NPU	~$2B (est.)
Intel	SambaNova	Pending	Boost inference silicon portfolio	Enterprise AI inference	Reconfigurable Dataflow	~$1.6B (rumored)
Marvell	Celestial AI	Announced Dec 2025	Own photonic IP for scale-out AI	Optical interconnect for AI clusters	Co-packaged optics	$3.25B upfront
Nvidia	Groq (IP)	Announced Dec 2025	License low-latency inference IP	Efficient inference for real-time AI	LPU + SRAM stack	Undisclosed

Slide 2: AI Compute Surge — Two Distinct Waves

Key Message: Training and inference are distinct workloads imposing different compute, memory, interconnect, and driving different system architectures.

Wave 1: Foundation Model Training

Led by: OpenAI, Google, Meta
Requires: 100K+ XPUs, interconnected via scale-up and scale-out
Characteristics: One long-running (~3 months) workload

Wave 2: Foundation Model Inference

Driven by Applications: Searching, Coding, Agents, Content Generation
Requires: Clusters of 100’s XPUs.
Characteristics: Billions of concurrent short-lived (~10 mins) sessions ``

Slide 3: AI Inference at Data Center Scale Is Exploding

Key Message:
Inference is now the dominant consumer of data center compute and is growing exponentially across applications.

Google reported 1.3 quadrillion tokens per month (~43 trillion per day), growing rapidly across search, code generation, agents, and enterprise workloads.
Inference is no longer a downstream by-product of training—it has become the primary compute workload.
Hyperscalers are re-architecting inference infrastructure away from general-purpose GPUs toward specialized, cost-optimized silicon.

Slide 4: Inference Is Commoditizing — TCO Replaces CUDA

Key Message: Inference is becoming a TCO game, not a TOPS contest.

Old Stack (Nvidia Moat):

CUDA + GPU + HBM + NVLink
High BOM, power inefficiency, vendor lock-in
New Stack:
RISC-V (Meta/Rivos, Qcom/Ventana, Tenstorrent) + NPU + DDR/SRAM + PCIe/UA/UE
Lower power, latency, and TCO
Design Shifts:
Domain-specific NPUs replace general-purpose GPUs
LPDDR (Qcom), GDDR (Tenstorrent), and on-chip SRAM (Groq) replace HBM
Tightly integrated CPU/NPU/memory inference silicon prioritizng power efficiency over peak TOPS.

Slide 5: Foundation Model Competition Drives Super-Scale Training

Key Message:
As the number of model leaders shrinks, the ability to scale training infrastructure becomes the core competitive edge.

The foundation model landscape is consolidating around a few global leaders (Google, OpenAI, Anthropic, xAI).
While developer count shrinks, training cost, model size, and system complexity are increasing.
Access to massive, reliable, scalable training infrastructure becomes a decisive advantage—not cost.

Slide 6: AI ASIC Vendors Shift Upmarket

Key Message:
As inference profit margins shrink, ASIC vendors are moving into training infrastructure where photonics become strategic.

Training requires:
- Massive compute scale-up, scale-out, and scale-across
- Ultra-high-bandwidth, low-latency interconnects
Photonic fabric is emerging as key enabler:
- Celestial AI (acquired by Marvell)
- Lightmatter (backed by Google)
- Ayar Labs (backed by Nvidia, AMD, Intel)
Photonic unlock the super-scale compute and memory pooling where next-gen AI system architectures is upon

Slide 7: Strategic Implications for MTK

Key Message:
MTK must act on two fronts: build photonic capability to compete in training infrastructure and capture inference opportunities

1. Invest in Photonic Technologies to Move Upmarket into Training and position as a relevant player for existing business

2. Evaluate Inference Server Opportunities

Potential customers, SAM, and impact on ASIC business
Assess and invest in server-class technologies: NPU (based on mobile NPU and scale-up required), CPU, system architecture, software