Insiders say the future of AI will be smaller and cheaper than you think

AI industry insiders forecast a shift from massive general-purpose LLMs toward smaller, task-specific AI agents that are cheaper to build and can run locally; firms cited include Superhuman, Mozilla and ARM. HSBC’s analysis highlights OpenAI’s large-scale capital challenge—the company claims $20 billion in revenue, has reportedly committed $1.4 trillion to data-center expansion, and even with potential $200 billion-plus revenues by 2030 would still need roughly $207 billion more funding—while investors such as IBM Ventures (a $500 million AI fund) are backing niche enterprise plays like Not Diamond’s model-router. The takeaway for allocators: structural opportunities in fit-for-purpose models and model-routing services, but sizeable financing risk remains for mega-scale LLM strategies.

Analysis

Market structure: Winners are edge/agent enablers (ARM, on-device inference silicon, niche SaaS with vertical data moats) and orchestration/software (model routers, app-store-like agent platforms); losers are scale‑dependent, capex‑heavy LLM plays (large portions of GOOGL/OpenAI economic thesis). If compact models routinely use 10–100x fewer parameters for many tasks, GPU-hour demand for those workloads could drop 30–60% over 24–36 months, shifting pricing power to software IP and domain data owners.

Risk assessment: Tail risks include regulatory clamps on data use or antitrust actions that break platform bundling (6–24 months), and a funding shock that forces consolidation of expensive LLM players (12–36 months). Hidden dependency: many “small” agents still route to giant pretrained models/APIs today, creating concentration risk in a few API providers; a failure/price shock at those providers would cascade to agent startups quickly.

Trade implications: Prefer long exposure to ARM (edge IP) and IBM (enterprise AI tooling/model-routing) with 6–12 month horizons and tactical hedges against hyperscaler weakness (GOOGL). Consider pair trades (long ARM, short GOOGL) and options: buy-call spreads on ARM or 9–12 month LEAPs; buy put spreads on GOOGL sized to 0.5–1% of portfolio to cap downside while collecting time decay.

Contrarian angles: Consensus underappreciates that base LLMs may remain indispensable for pretraining and transfer learning—meaning hyperscalers retain a backstop revenue stream and Nvidia-like hardware winners may be underowned. Also, fragmentation raises integration/security costs that benefit established enterprise vendors (IBM) and could slow pure-play agent monetization, creating opportunities to sell early exuberance and buy durable incumbents.

Key Decisions for Investors

• Establish a 2–3% long position in ARM (ticker: ARM) over 6–12 months; if implied volatility is high, execute a 6–9 month call spread to target 20–40% upside while limiting premium spent. Entry: on pullback >8% or after next quarterly beat; Exit: +25–35% or at 12 months.
• Initiate a pair trade: long ARM 1.5% / short GOOGL 1.5% (hedged notional) to capture rotation from hyperscalers to edge/agent enablers. Timeframe: 3–12 months; stop-loss if spread widens against position by 12% intraday.
• Allocate 1–2% to IBM (IBM) equity for enterprise model-routing and hybrid deployment revenues with a 12‑month horizon; target +15% upside. Reduce exposure if IBM misses revenue guidance or if announced clients for model-router technology fall below 2 major wins in 90 days.
• Buy a 3–6 month put spread on GOOGL (size 0.5–1% notional, strikes ~5–15% OTM depending on pricing) as downside protection against capex re-rating for hyperscalers; roll or close if drawdown in GOOGL >20% or market confirms durable rotation to edge agents.