Compute, hardware innovation, model performance, and embodied/agentic deployments

AI Infrastructure & Agentic Research

The AI compute and hardware landscape for 2026–2028 is rapidly evolving, driven by multipolar geopolitical dynamics, strategic hardware-model co-design, and breakthroughs in model performance and agentic AI deployments. Recent developments reinforce and deepen trends identified earlier, underscoring how AI infrastructure is becoming increasingly complex, distributed, and performance-optimized to meet the demands of embodied and autonomous systems at scale.

Multipolar Compute and Hardware-Model Co-Design: Expanding Alliances Amid Geopolitical Fragmentation

The shift from a U.S.-centric compute supply chain to a multipolar architecture is accelerating, with strategic partnerships and geopolitical maneuvering reshaping AI hardware and model ecosystems:

DeepSeek’s Model Withholding Remains a Symbol of Tech Sovereignty:
DeepSeek’s continued decision to withhold its flagship AI models from U.S.-based chipmakers, instead favoring partners like Huawei, highlights the ongoing geopolitical tensions and the growing push for technology sovereignty. This move is forcing hyperscalers and silicon vendors alike to diversify supply chains, localize infrastructure, and develop alternative co-design pipelines outside traditional U.S. ecosystems.
AMD–Meta and Intel–SambaNova Partnerships Deepen:
The AMD–Meta collaboration is intensifying efforts to co-develop domain-specific accelerators optimized for Meta’s portfolio of agentic and embodied AI applications. This alliance is a direct challenge to Nvidia’s dominance, with AMD leveraging Meta’s scale and AI expertise to tailor silicon innovation.
Meanwhile, Intel’s $350 million investment in SambaNova Systems following failed acquisition talks signals Intel’s strategic commitment to domain-specific AI silicon, particularly for compliance-heavy and mission-critical workloads. SambaNova’s technology complements Intel’s evolving data center strategy, positioning it to better compete in heterogeneous AI compute markets.
Silicon Startup Funding and Acquisitions Expand Options:
- MatX, having secured $500 million in funding, is progressing aggressively on model-to-silicon AI chips that promise up to 10x inference performance improvements by hardwiring agentic AI pathways.
- Taalas continues to claim pioneering throughput records, with token processing speeds reaching 17,000 tokens per second, challenging GPU-centric paradigms.
- Nvidia’s $60 million acquisition of Israeli startup Illumex advances its software-hardware integration stack, emphasizing runtime efficiency and tighter coupling of AI models with underlying silicon.
- New entrants like Callosum are also drawing investor interest, further diversifying AI chip innovation.
Hyperscalers Expand Sovereign and Private Cloud Infrastructure:
Microsoft is expanding sovereign cloud capacity aggressively in emerging markets and the Middle East, embedding secure AI services that meet strict data residency and compliance demands. Google Cloud, in partnership with DeepMind, is pushing forward in hardware-model co-design to optimize heterogeneous compute resources for next-generation AI workloads.
Notably, NTT DATA and Ericsson have teamed up to scale private 5G and physical AI solutions for enterprises, enabling edge compute environments critical for latency-sensitive embodied AI deployments in manufacturing, logistics, and healthcare.

Hardware Innovation: Photonics, AI-Driven SoC Verification, and Model-to-Silicon Breakthroughs

Hardware innovation is not only scaling performance but also improving efficiency and sustainability through novel technologies:

Photonics Chips Gain Traction:
Apple’s acquisition of photonics AI startup invrs.io continues to signal a strategic pivot toward light-based AI chips. Photonic processors use optical signals instead of electrons, delivering orders-of-magnitude improvements in energy efficiency and throughput. These chips are especially critical for reducing AI’s growing carbon footprint and enabling compute-intensive workloads at the edge.
AI-Driven SoC Verification Enables Faster Silicon Development:
A new AI-driven SoC verification flow, developed through a partnership between Breker Verification Systems and Moores LabAI, is revolutionizing silicon design cycles. This flow integrates AI to automate complex verification tasks, reducing time-to-market and improving reliability of AI chips tailored for agentic and embodied AI workloads.
Model-to-Silicon Advances and Ultra-Low-Latency Models:
Startups like BeyondMath continue pioneering physics-based simulation and embedding LLM architectures directly into silicon, which dramatically reduces inference latency and power consumption, supporting real-time edge and industrial use cases.
The release of ultra-fast language models such as Mercury 2, boasting generation speeds 13x faster than Anthropic’s Claude Haiku by breaking traditional sequential token generation, is complemented by OpenAI’s gpt-realtime-1.5 model. The latter offers enhanced instruction adherence and improved speech interaction capabilities, powering robust voice workflows in real time.
Memory Primitives Improve Agentic AI Capabilities:
The introduction of DeltaMemory, touted as the fastest cognitive memory for AI agents, addresses a crucial bottleneck: agent forgetfulness between sessions. By enabling long-term, persistent memory with rapid recall and update speeds, DeltaMemory empowers agents with enhanced contextual awareness and continuity, vital for embodied and autonomous AI applications.
Multimodal and Multilingual Models Expand Edge and Sovereign AI:
Models like Alibaba’s Qwen 3.5 Medium continue to push local inference performance to near commercial-grade levels, enabling sovereign and offline AI deployments. Research in test-time training (TTT) with key-value binding methods further improves dynamic model adaptation without costly retraining, enhancing agility in production environments.
Vision-language-action models such as VLANeXt improve embodied AI’s contextual intelligence, enabling more nuanced decision-making for robots and autonomous systems.

Orchestration Platforms and Observability Tools Mature for Production-Grade Agentic AI Deployments

The complexity of deploying multi-agent and embodied AI systems at scale is driving a new generation of development and runtime tooling:

Agent Orchestration Platforms Expand:
Platforms like Union.ai (recently closed a $38.1 million Series A), Tensorlake AgentRuntime, and AWS Strands Labs are providing scalable, distributed environments for building, deploying, and managing multi-agent AI and robotics applications in production. These platforms emphasize governance-first design, observability, and compliance, addressing critical enterprise and regulatory demands.
New Entrants and Marketplaces Democratize Agent Creation:
The launch of Perplexity’s ‘Computer’ AI agent, which coordinates 19 underlying models and is priced at $200 per month, marks a significant step in agent orchestration and multi-model integration for businesses and developers. Alongside Microsoft’s Copilot Studio and Agent Framework and Google’s Opal, these ecosystems empower no-code and low-code creation of autonomous agents, accelerating innovation and adoption across industries.
Observability and Compliance Tools Gain Enterprise Traction:
Tools such as OpenTelemetry and New Relic’s AI Agent Platform are delivering comprehensive telemetry capture, fault detection, audit trails, and compliance monitoring essential for mission-critical AI systems. These capabilities enable enterprises to maintain operational resilience, meet stringent regulatory requirements, and build trust in AI deployments.

Edge, Private 5G, and Sovereign Cloud Scale Agentic AI in Regulated and Latency-Sensitive Domains

The deployment of embodied AI and agents is expanding beyond centralized clouds into edge and sovereign environments:

Private 5G and Edge AI for Enterprises:
The NTT DATA–Ericsson partnership is advancing private 5G and physical AI networks that support low-latency, high-reliability AI applications in manufacturing, logistics, and healthcare. These networks provide the necessary infrastructure for real-time embodied AI and autonomous systems operating at the edge.
Sovereign Cloud Expansions:
Microsoft and Google Cloud continue to grow sovereign cloud offerings, embedding AI services with strict data residency, privacy, and compliance controls tailored for regulated sectors such as defense, finance, and healthcare. These expansions are critical for enterprises navigating complex regulatory landscapes while adopting advanced AI.
On-Device AI Agent Deployments:
Consumer and industrial devices increasingly incorporate on-device AI agents for privacy-preserving, low-latency interactions. The Samsung Galaxy S26 series integrates Perplexity AI, while startups like Mirai are innovating to improve on-device inference performance, meeting growing demand for autonomous AI capabilities at the edge.

Validated ROI and Governance Drive Enterprise Adoption Across Key Verticals

Agentic AI is moving decisively from experimental to commercial-scale deployments, with measurable returns and governance frameworks enabling trust and compliance:

Mobility and Robotics:
The Wayve–Uber $2.5 billion robotaxi partnership continues scaling, integrating autonomous fleets with gig economy platforms. Advances in zero-shot dexterous manipulation, exemplified by policies like SimToolReal, empower robots to execute complex, flexible tasks without retraining, critical for logistics and manufacturing automation.
US shipyards are trialing AI-driven uncrewed shipbuilding, employing embodied AI for precision welding and assembly, showcasing AI’s potential in heavy industry.
Manufacturing, Retail, and Defense:
Enterprises report up to 3x sales conversion improvements driven by AI chatbots in retail, alongside workflow optimization and quality control gains in manufacturing. Defense agencies are streamlining procurement and maintenance through agentic AI, reducing operational costs and timelines.
Governance-first observability platforms, integrating telemetry and policy controls, are essential in mitigating operational risks and meeting compliance mandates, fostering broader enterprise confidence and AI adoption.
Governance as a Strategic Imperative:
Integrated governance and observability are no longer optional but foundational. AI deployments increasingly embed real-time monitoring, fault tolerance, and auditability to satisfy regulatory scrutiny and operational continuity, especially in mission-critical environments.

Conclusion

The AI compute and hardware ecosystem in 2026–2028 is defined by multipolar geopolitical fragmentation, strategic hardware-model co-design, and breakthrough innovations in photonics, model-to-silicon integration, and ultra-fast large language models. These advances underpin a new generation of agentic and embodied AI systems deployed across cloud, edge, and on-device environments.

Emerging partnerships (AMD–Meta, Intel–SambaNova), startup innovations (MatX, Taalas, Illumex), and hyperscaler initiatives (Microsoft, Google Cloud, NTT DATA–Ericsson) are collectively diversifying the silicon and software landscape, enabling scalable, secure, and sovereign AI infrastructures.

Simultaneously, orchestration platforms like Union.ai, Perplexity Computer, and AWS Strands Labs are maturing to support complex multi-agent and robotics deployments with integrated governance and observability, critical for enterprise and regulated industries.

As one industry analyst summarized:

“The AI compute battlefield is no longer just about raw power. It’s a complex dance of geopolitical strategy, hardware-software co-innovation, and trustworthy governance—paving the way for agentic AI to become the backbone of future enterprise and industrial transformation.”

This evolving landscape signals that the next frontier of AI will be defined by resilience, integration, and trustworthiness, not just performance benchmarks.

References to New Developments:

Perplexity’s launch of the “Computer” AI agent coordinating 19 models at $200/month
NTT DATA and Ericsson’s private 5G and physical AI enterprise deployment partnership
Breker Verification Systems and Moores LabAI’s AI-driven SoC verification solution
OpenAI’s gpt-realtime-1.5 model enhancing real-time voice workflows
DeltaMemory’s breakthrough in persistent cognitive memory for AI agents

This comprehensive synthesis highlights the shaping forces and technologies propelling AI from raw compute arms races toward integrated, sovereign, and production-grade systems that enable the agentic AI era.

Sources (519)