In “The Second Half,” Yitan Liu masterfully dissects AI’s trajectory, contrasting the “first half”—dominated by groundbreaking models like Transformers, which amassed over 160,000 citations while datasets faded into the background—with the emerging “second half.” Here, the focus pivots from innovating training techniques to pinpointing real-world problems and crafting evaluations that measure genuine impact.

Liu highlights the “utility problem”: AI has demolished benchmarks in chess, Go, SATs, and even Olympic math, yet its footprint on global GDP remains negligible. The antidote? Revolutionary evaluations incorporating human-in-the-loop dynamics, sequential tasks, and long-term memory to bridge lab triumphs with economic revolutions.

But Liu’s vision peaks with a speculative powerhouse: imagine “Macrohard,” a company leveraging AI agents to craft any software bespoke to user needs.

This isn’t mere automation—it’s a paradigm where AI doesn’t just assist but entirely constructs tailored applications on demand. As we hurtle into this second half, trillion-parameter models, with training costs soaring into billions, will fuel such entities.

Yet, the real earthquake? Elon Musk’s xAI with their Macrohard’s project could eviscerate today’s tech behemoths like FAANG (Facebook/Meta, Amazon, Apple, Netflix, Google/Alphabet) and SaaS stalwarts, rendering their one-size-fits-all empires obsolete. Let’s speculate on where and how these AI juggernauts will materialize, and why the software landscape may never recover.

Macrohard Unleashed: The Bespoke Software Revolution and Its Assault on SaaS Empires

At its core, Macrohard embodies the second half’s promise: AI agents that ingest requirements and output fully functional, customized software. No more rigid platforms; instead, dynamic, adaptive code sculpted for precise business needs.

This upends the SaaS model, where companies like Salesforce, Adobe, or Zoom peddle subscription-based tools that users mold to fit—often imperfectly.

AI agents, empowered by vast language models, could automate code generation, testing, and deployment, slashing development timelines from months to minutes and democratizing bespoke solutions.

Consider the ripple effects on FAANG and beyond:

• Eroding the SaaS Subscription Fortress: Traditional SaaS thrives on recurring revenue from generalized apps—think Microsoft’s Office suite or Google’s Workspace. But with AI agents, businesses could summon bespoke alternatives: a custom CRM rivaling Salesforce, tailored to niche workflows without the bloat. Microsoft’s own Charles Lamanna predicts AI “business agents” will obsolete static apps by 2030, dynamically adapting to user contexts and eroding the need for monolithic platforms. This “agentic AI” shifts from passive tools to proactive entities, potentially halving SaaS market share as enterprises opt for on-demand, cost-effective custom builds.
• Dismantling FAANG’s Data Moats and Ecosystems: Giants like Amazon (AWS) and Google dominate via ecosystems locking users in. Macrohard-style AI disrupts this by enabling seamless, interoperable software creation without vendor lock-in. For instance, instead of relying on Apple’s App Store or Netflix’s algorithms, AI could generate personalized apps or content pipelines, fragmenting their monopolies. Startups armed with AI could infiltrate horizontal systems like HR (challenging Workday) or e-commerce (undercutting Shopify), exposing vulnerabilities in legacy SoRs (systems of record). As PwC’s 2025 predictions note, AI’s role in business transformation will favor agile disruptors over entrenched players.
• Accelerating the ‘Buy vs. Build’ Flip: Historically, high costs favored buying off-the-shelf SaaS. AI flips this: tools like Replit or Bolt lower barriers, making custom software viable for SMBs. In 2025, McKinsey reports software firms embedding agentic AI into products, but this could backfire—empowering users to bypass them entirely. Experts are split: some see AI “eating” SaaS, others resilience, but the consensus leans toward hybrid disruption where AI agents handle complex, sequential tasks humans once managed.

This isn’t hyperbole; Bessemer Venture Partners’ State of AI 2025 flags opportunities for startups to disrupt via AI-exposed gaps in enterprise software. Yet, incumbents aren’t idle—Salesforce’s Agentforce exemplifies defensive plays, but if Macrohard materializes, it could commoditize software creation, tanking valuations and forcing pivots.

Forging the Disruptors: Where and How Second-Half AIs Will Be Built

To birth Macrohard’s agents, we’ll need colossal infrastructures for trillion-parameter models, with aggregate costs eclipsing trillions. Training Meta’s Llama 4, at 2 trillion parameters, demands 520 trillion TFLOPs—harbingers of the scale ahead.

• Geographic Hotspots: The US leads, with Virginia’s Data Center Alley and Texas’ energy deregulation drawing billions from Microsoft ($80B) and Meta. Middle Eastern powerhouses like Saudi Arabia ($600B investments) and UAE ($1.4T) offer cheap, renewable energy for solar-powered mega-centers. Asia (China’s state-backed clusters) and Europe (green facilities in Sweden) round out the map, potentially creating a fragmented “AI grid” amid geopolitics.
• Construction Blueprints: Hyperscale clusters of NVIDIA B200 GPUs, modular “AI pods,” and nuclear-powered sites like OpenAI’s 4.5GW Stargate tackle energy hogs (536 TWh globally in 2025). Collaborations fuse hyperscalers with sovereign funds, while AI-optimized designs cut power by 50%. McKinsey’s $6.7T forecast by 2030 underscores the industrial shift.

These builds enable iterative evaluations—simulating bespoke software creation at scale—propelling Macrohard’s disruption across EVERY software vertical.

In Liu’s halftime analogy, the second half isn’t just evolution; it’s revolution.

Revolutionary projects like Macrohard’s bespoke AI could shatter FAANG’s dominance, birthing a world of fluid, intelligent software built by everyone.

The new moats might not be moats but creativity with energy, intelligence and constant evolution

As 2025 unfolds, even the once seemingly unbeatable titans will fall and be replaced by something entirely different and new.

Perhaps they’ll be tiny human employee companies with millions of bespoke agentic employees adding value through intelligence everywhere and in every application from software through to physical infrastructure and robotic labor.

Related Reading

• The AI-Biology Flywheel: Why Intelligence and Biotechnology Are Accelerating Each Other
• AlphaFold and Drug Discovery: How AI Is Finding Molecules That Extend Lifespan
• AI for Science: How Artificial Intelligence is Revolutionizing Drug Discovery, Biology & Research
• The Evolution of Layer-1 Blockchains and Smart Contracts: Ethereum, Solana, and Beyond