May Mobility's $1.4B SPAC puts pure-play robotaxis on the Nasdaq
Robotaxi economics finally get a public scorecard. Elsewhere: a 4B model that out-plans Postgres, and Apple's agents for the least glamorous job in enterprise AI.
May Mobility is going public through a SPAC merger at a $1.4 billion pro forma enterprise value, and says the combined company will be the first US-listed business focused entirely on autonomous ride-hailing. The Michigan company is merging with ACP Holdings Acquisition Corp, a blank-check vehicle set up by Houston investment manager Atlas Credit Partners, in a deal that pairs a $120 million private investment in public equity with up to $217 million held in the SPAC's trust account — a maximum of about $337 million in gross proceeds, before any shareholder redemptions. May Mobility runs autonomous Toyota Siennas in three US locations, carries a Lyft partnership in Atlanta, gives rides in Eden Prairie and Grand Rapids, Minnesota, and recently started a first trial deployment in Japan, with a commercial launch in Arlington, Texas, alongside Uber targeted for late this year or early 2027.
The business model is the part worth watching. May Mobility does not own its robotaxis; it sells vehicles to fleet partners over time and keeps control of remote supervision and software updates, collecting fixed fees or per-trip licensing fees. That "asset-light, partnership-first" pitch is a direct contrast with Waymo, which owns and operates its own fleet. The financials disclosed alongside the listing are less flattering: roughly $10 million in revenue last year against a cash burn of about $93 million, with more than 550,000 paid rides and over a million miles logged to date. Proceeds are earmarked for research and development — especially removing safety drivers — plus supply-chain work to cut bill-of-materials costs. Our read: the SPAC answers a question the private rounds never had to. Robotaxi optimism has been priced by venture investors for a decade; now it gets priced by public shareholders who can see a nine-to-one burn-to-revenue ratio, and who can redeem rather than wait. Redemptions are the real number in this deal — $337 million is a ceiling, not a floor.
A 4B open-weights model post-trained with supervised fine-tuning plus agentic reinforcement learning cut query planning time versus Postgres's own planner by 1.81x. The experiment, published by engineer Rohan Bansal and circulating on Hacker News, takes a small Qwen model and teaches it to emit planner hints rather than answer questions: the model proposes candidate plans for a query, Postgres measures each one against its default plan, and the measured speedup becomes the training reward. Training used a LoRA adapter on two consumer RTX 3090s with rented H100s for the longer runs, and an adapted GRPO objective that anchors each rollout's advantage to the default plan's execution time instead of the batch's own rewards. Taking the best feedback across three rollouts — up to 15 candidate plans per query — the final checkpoint hit 1.81x on both geometric-mean and total-workload speedup across the 113-query Join Order Benchmark.
The caveat is the same one that applies to every learned optimizer: the win is amortized, not general. Beating Postgres on a one-off query is not the claim — the claim is that for analytic queries run thousands of times, spending tens or hundreds of executions upfront on training and search pays back. That reframes small models as a per-database asset rather than a general-purpose substitute, and it is the most concrete demonstration yet that a two-person team with consumer GPUs can train against a domain-specific environment and beat a 30-year-old piece of infrastructure on its home turf.
Apple detailed Glyph, a production agent system that writes column descriptions and sensitivity labels for enterprise data catalogs — the documentation debt that quietly undermines data governance. Glyph runs two cooperating agents as stateful graphs: a descriptor that grounds each column's description in the pipeline source code that produces it, retrieved on demand from an internal GitHub through a tool loop, and a tagger that assigns labels from a governed ontology of 275 leaf categories by running three strategies in parallel — description-based, regex-based, and a metadata encoder over a vector store — then fusing their ranked outputs. Apple fine-tuned a six-layer MiniLM encoder with an in-batch contrastive objective, lifting same-tag retrieval from NDCG@10 0.55 to 0.92, and reports end-to-end tagging quality under a recall-weighted F2 objective with per-tag provenance and graceful degradation. The value-free design is the interesting choice: Glyph classifies columns without reading the data in them, which is what makes it deployable in regulated environments where a scanner that samples values is a non-starter.
What to watch: whether May Mobility's public comps hold through redemption, and whether the query-planning result gets reproduced on a production database rather than a benchmark.
Would you buy a robotaxi pure-play at $1.4 billion before it has shown a path past a $93 million burn — or wait for the first profitable quarter? Tell us in the comments.
Sources: TechCrunch · Reuters · Axios · Rohan Bansal — QORL · Hacker News discussion · Apple Machine Learning Research · arXiv:2609.10430
