SpaceXAI launches Grok 4.5, its first built with Cursor's help

SpaceXAI (rebranded from xAI) launched Grok 4.5, its default AI coding-agent model trained on tens of thousands of NVIDIA GB300 GPUs and priced at $2/M input tokens and $6/M output tokens (vs OpenAI’s $5/$30 for the most powerful tier). The model is positioned as faster and lower-cost than “flash” models and able to generate functional apps from minimal prompts. The launch follows a Cursor partnership with potential for a $10B investment or a $60B acquisition later in 2026, while access is restricted in the EU until mid-July.

Analysis

This is more a pricing-war signal than a single-product launch. If a frontier lab is willing to undercut premium inference pricing while claiming top-tier quality, the near-term beneficiary is compute demand: cheaper tokens usually expand usage faster than they destroy margin, which is constructive for NVDA over the next 1-3 months as hyperscalers and model vendors keep buying training and inference capacity.

The second-order loser set is the AI application layer whose pricing assumes scarce, expensive tokens. Expect enterprise procurement to push for repricing at the next renewal cycle, especially for copilots, code assistants, and document automation tools that are already vulnerable to feature parity. That creates a subtle squeeze: model vendors may cut price to win share while app vendors lose ARPU, so the market could be misreading this as uniformly bullish for AI when it is really a margin transfer upstream to hardware and downstream to customers.

The contrarian risk is that benchmark wins do not equal durable monetization. If usage does not re-accelerate, lower token prices simply compress gross margin and force a capex arms race, which would be negative for the broader AI stack in 6-18 months. For NVDA, the thesis breaks if GB300 deployment or hyperscaler capex commentary softens; for software shorts, it breaks if management teams report no meaningful AI pricing pressure in the next two quarters.

Key Decisions for Investors

• Long NVDA on any 2-4% post-launch pullback; use a 1-2 month call spread to express the view that lower model prices expand usage faster than they dent GPU demand. Risk/reward improves ahead of the next hyperscaler capex read-through.
• Pair trade: long NVDA / short a diversified AI software basket (e.g., MSFT, NOW, CRM, SNOW) for 1-3 months to isolate the cost-curve beneficiary versus the margin-compressed application layer. Falsify if enterprise AI spend broadens without pricing pressure commentary.
• If LUNA maps to a public AI monetization proxy, treat it as a watch item rather than a fresh long: the launch increases the odds that low-end model economics get commoditized first. Wait for evidence of usage growth before paying for multiple expansion.
• Set an alert on NVDA and hyperscaler capex commentary into the next earnings cycle; if order growth or delivery timing slips, take profits quickly because the market is likely front-running a stronger infrastructure cycle than the data currently supports.