Issue #7
March 14 – 20, 2026
~9 min read

The Trillion-Dollar Inference Thesis

Nvidia's GTC 2026 dominated the week, reframing AI as an industrial production system where inference — not training — is the dominant workload. Jensen Huang doubled the company's demand projection to $1 trillion through 2027, unveiled the Vera Rubin platform, the Groq LP30 inference chip, and NemoClaw, an enterprise agent operating system. Meta signed a $27 billion five-year deal with Nebius for Vera Rubin capacity. Meanwhile, the Microsoft-OpenAI partnership fractured further, and the attack surface widened from Cisco zero-days to iPhone spyware campaigns targeting Southeast Asia.
Contents
Singapore Tech News This Week
KKR-Singtel consortium's STT GDC acquisition at S$13.8B enterprise value moves toward H2 2026 close
DayOne Data Centers nears confidential U.S. IPO filing
Miro announces Asia hub in Singapore, targeting SEA, India, and South Korea
I. GTC Rewrites the AI Roadmap Around Inference — 1.5 min read
Inference, Not Training, Is the Dominant Workload
Bottom line: Nvidia's GTC declared inference — not training — as the dominant AI workload, and backed the claim with $1 trillion in demand, a new inference chip, and an agent OS.

GTC 2026 reframed the AI industry around a single thesis: the future is inference at industrial scale. Jensen Huang doubled Nvidia's demand projection to $1 trillion through 2027, up from $500 billion at GTC 2025, driven by a shift from model training toward inference, robotics, and agentic AI. The Vera Rubin platform — packaging seven chip types into five rack-scale computers — entered production alongside Blackwell GB300.

Nvidia also unveiled the Groq LP30, its first inference-optimised chip from the ~$17–20 billion Groq licensing deal, with 500MB on-chip SRAM designed to complement GPUs. And NemoClaw, an enterprise-secure agent operating system, positions Nvidia as the platform layer for agentic AI — not just the hardware supplier.

The capital commitments matched the ambition. Meta signed a $27 billion five-year deal with Nebius for dedicated Vera Rubin capacity — one of the largest AI infrastructure contracts announced to date. With Micron reporting Q2 revenue of $23.86 billion on surging AI memory demand and Alibaba reportedly raising AI computing prices by up to 34%, the supply side is confirming what GTC declared: AI infrastructure demand is real, but the industrial cost of meeting it is growing enormous.

II. Cloud Partnerships Fracture Under Commercial Pressure — 1.5 min read
The Exclusive-Partnership Era Is Ending
Bottom line: Microsoft is threatening to sue over a $50B Amazon–OpenAI deal, while OpenAI simultaneously signs with AWS for government work — the exclusive-partnership era in AI cloud is ending.

The most consequential AI partnership of the past five years is fracturing over cloud exclusivity. Microsoft is weighing legal action against OpenAI and Amazon over a $50 billion deal that makes AWS the exclusive third-party cloud provider for OpenAI's Frontier enterprise platform. Microsoft argues this breaches its Azure exclusivity agreement.

Separately, OpenAI signed a new AWS deal to sell AI tools to U.S. government customers across classified and unclassified work — a direct expansion into government territory that was previously Microsoft's domain. The Anthropic-Pentagon dispute also escalated, with industry groups filing an amicus brief urging a pause on the supply-chain risk designation ahead of a March 24 hearing.

The two disputes share a common thread: the era of exclusive AI partnerships is giving way to a more fragmented, multi-vendor landscape. The question is whether the legal architecture governing these relationships can keep pace with the commercial pressure to break them.

III. Trusted Tools Become Attack Vectors — 1.5 min read
The Attack Surface Is the Management Plane
Bottom line: Attackers weaponised a Cisco zero-day since January, turned Microsoft Intune against its own users, and compromised iPhones across four countries — the common thread is exploitation of infrastructure organisations already trust.

This week's threat landscape converged on a single pattern: attackers exploiting the tools and infrastructure that organisations already trust. A Cisco Secure Firewall Management Center zero-day (CVE-2026-20131, CVSS 10.0) was disclosed — allowing unauthenticated remote code execution as root. Amazon's MadPot sensor network detected exploitation since January 26, more than a month before public disclosure.

Separately, Iran-aligned Handala used Microsoft Intune — an organisation's own mobile device management platform — to issue remote wipe commands at medical manufacturer Stryker. And DarkSword, a spyware platform embedded across Ukrainian websites, was discovered targeting iPhones running vulnerable iOS 18, with campaigns extending to Saudi Arabia, Turkey, and Malaysia.

The pattern extends to AI infrastructure itself. Oasis Security researchers disclosed three chained vulnerabilities in Claude that could steal user chat data without malware. The common denominator: the attack surface is no longer the perimeter — it is the management plane, the device trust layer, and the AI tools themselves.

NUS Connection

NUS SoC Asst. Prof. Trevor Carlson's group published SSBleed (HPCA 2026), demonstrating the first non-speculative side-channel attack on Armv9 CPUs via the Memory Dependence Predictor — showing that even next-generation chip architectures carry undocumented hardware vulnerabilities. As AI inference moves to Armv9-based edge devices, the attack surface extends from software to silicon.

IV. The Capex Reckoning Meets Physical Buildout — 1.5 min read
Markets Begin Pricing Infrastructure Risk
Bottom line: Investor confidence is starting to separate AI growth stories from AI capex risk — and organised labour is beginning to push back on the deployment side.

Micron beat revenue expectations with $23.86 billion, then said it would increase 2026 capital spending by $5 billion to over $25 billion — and its stock fell. Markets are beginning to price infrastructure credibility separately from growth narratives.

The physical buildout continues regardless. Germany announced plans to double domestic data centre capacity. In Southeast Asia, Johor is projected to approach 1GW of data centre capacity by end-2026. Meanwhile, approximately 2,400 Kaiser mental health workers struck over fears that AI will displace therapists — the first major organised labour action explicitly targeting AI deployment in healthcare.

V. Research Highlights — 4 min read
Recent Papers Ranked by Relevance
1. "LLMs + Security = Trouble"
Livshits (Imperial College London) · arXiv, Feb 2026
Argues that AI-generated code introduces security vulnerabilities faster than AI-powered detection tools can find them, proposing constrained decoding during generation as a more promising approach than post-hoc repair.
Relevant as GTC's NemoClaw positions Nvidia as the agent OS layer — if code those agents generate is insecure by default, the platform inherits vulnerability at scale.
2. "Securing the AI Supply Chain"
Nguyen, Le & Babar (University of Adelaide) · ICSE 2026
Empirically maps developer-reported security issues across the AI supply chain — from model poisoning to malicious packages. The Cisco FMC zero-day and Handala's Intune weaponisation demonstrate these attacks are no longer theoretical.
3. "SSBleed: Non-Speculative Side-Channel Attacks on Armv9 CPUs"
Liu, Zheng, Zhang, Ju, Wang, Zhang & Carlson (NUS SoC / Tsinghua) · HPCA 2026
Demonstrates the first non-speculative data leakage via the Memory Dependence Predictor on Armv9 performance cores. Existing mitigations are insufficient.
As Nvidia's Vera Rubin and Groq LP30 drive inference to new hardware, undocumented predictor behaviours in next-gen CPUs represent an attack surface current defences don't address.
4. "Labor Market Impacts of AI: A New Measure"
Massenkoff & McCrory (Anthropic Research) · Mar 2026
Constructs a task-level AI exposure measure and finds limited evidence that AI has affected employment to date — though hiring of younger workers in exposed occupations shows early signs of slowing.
5. "Resolving the Automation Paradox"
Autor & Kausik (MIT) · arXiv, Jan 2026
Proves that a declining labor share driven by automation can paradoxically raise average wages — making the distributional question more urgent than the headline employment numbers suggest.
6. "ICT for Carbon Sinks"
Dong & Wu (Guilin U. / U. Chile) · arXiv, Mar 2026
Surveys how AI-driven approaches show measurable cost reductions in carbon capture. As Germany quadruples AI processing and Johor approaches 1GW, environmental cost becomes a deployment condition.
VI. Food for Thought — 1 min read
Three Things to Sit With

GTC's central claim — that inference has overtaken training — has infrastructure implications beyond Nvidia.

If tokens are the output of an industrial production system, then the entire stack needs redesigning around continuous serving, not batch training. Every data centre contract signed this week — Meta's $27 billion, Germany's doubling plan, Johor's 1GW target — is a bet on inference economics that have not yet been proven at this scale.

Microsoft's willingness to threaten litigation marks the end of the exclusive-partnership era.

OpenAI is now on AWS, Oracle, and Azure simultaneously. Anthropic works with Amazon, Google, and Microsoft for non-defence applications. The shift from exclusive to multi-vendor means cloud margins will compress, but no single provider can be a chokepoint.

The convergence of Cisco's zero-day, Handala's Intune weaponisation, and the Claude vulnerability chain suggests attacks will target the management plane — not the application layer.

Firewalls, device management platforms, and AI assistant APIs are all trusted infrastructure. When attackers compromise the tools that enforce security, the entire trust model inverts. As AI agents gain enterprise access via platforms like NemoClaw, every agent permission becomes a potential attack surface.

Back to Homehorizons.comp.nus.edu.sg