What MiniMax's stock surge-then-crash means for ad tech

The impact of MiniMax's AI stock surge on ad tech is easier to read in August 2026 than it was in March. The surge is no longer a live market story. It was a completed run from a Jan. 9 IPO price of HK$165 to a debut close near HK$345, then to a Mar. 18 peak close of HK$1,238, with an intraday high around HK$1,330. By Jul. 17 the stock was near HK$216, and by Jul. 30 it was around HK$203.80.[1]

That sequence matters because a 400% climb followed by a roughly 80% collapse can look, from a distance, like a verdict on AI advertising. It was not clean enough for that. A stock chart can show scarcity, momentum, and investor appetite; it cannot tell a media buyer whether an AI video model will improve Advantage+, Performance Max, or AI Max results next Monday.

The stock move ran ahead of the business

The cleaner read is financial, not creative. At the peak, MiniMax was associated with a market capitalization near HK$390 billion and a price-to-sales ratio above 700x. Its FY2025 revenue was US$79.0 million, up 158.9% year over year, while net loss was US$1.87 billion and adjusted net loss was US$250.9 million.[2]

Those numbers do not say the company had no product signal. They say the public valuation was not being carried by operating scale. The more plausible explanation is a free-float scarcity premium layered on top of AI enthusiasm. That is a trading condition, not an ad-account instruction.

For a growth team, that distinction is not academic. If the market cap had reflected durable revenue traction from advertisers, it might have been worth asking whether ad budgets were already shifting toward a new creative infrastructure. But a valuation that far ahead of revenue and losses is too noisy to treat as evidence that AI video tooling has been validated for performance marketing.

The useful signal is underneath the equity story

The MiniMax stock event is a bad guide for vendor selection. It is a useful excuse to look again at the economics of video generation.

Hailuo 2.3 Fast claims batch costs up to 50% lower, while Hailuo 02 has been positioned around roughly US$0.045 to US$0.08 per second as a budget silent-b-roll tier.[3] Treat those as vendor disclosures, not independent proof of ad performance. Still, the direction is hard to ignore: generating more raw video has become much cheaper.

That changes the creative supply equation. In many paid social and paid search accounts, the old constraint was the number of viable video cuts a team could afford to produce. If a buyer needed five new hooks, three product angles, two seasonal overlays, and several aspect-ratio versions, production bandwidth often became the reason the test never happened. Cheap generation weakens that excuse.

It does not remove the rest of the job. A raw model output is footage. It is not a finished Meta ad, YouTube Shorts cut, Performance Max asset set, or AI Max creative unit. Someone still has to write the script, decide what the clip is supposed to prove, reject bad motion and visual artifacts, add captions, handle voice quality, check whether the asset is safe to use, and package the result for the platform.

Cheap generation is not the same as easy finished ads

The practical mistake is to count generated clips as creative tests. They are not the same unit.

A media buyer cannot learn much from a folder of 200 loosely related clips. A testable ad needs a hypothesis. Maybe the hypothesis is that a product-in-use opening beats a founder-led opening. Maybe it is that silent b-roll with strong captioning can replace creator footage for a lower-intent audience. Maybe it is that a price-led hook will fatigue faster than a problem-led hook. Without that structure, more video just creates more review work.

The assembly layer is where a lot of the promised speed either appears or disappears. The clip has to be cropped, paced, captioned, paired with a voice or left silent intentionally, matched to a landing-page claim, exported in the right ratios, and named so results can be read later. If the team skips that discipline, the platform may still spend, but the buyer will not know what actually won.

This is especially true inside automated campaign types. Advantage+, Performance Max, and AI Max can find combinations that a buyer did not manually predict, but they do not absolve the team from feeding them coherent inputs. If ten assets differ by hook, offer, format, voice, product angle, caption style, and visual setting all at once, the learning is muddy even when the platform reports a winner.

The bottleneck has moved

When video generation was expensive, the bottleneck sat near production volume. When generation becomes cheap, the bottleneck moves downstream. The scarce work becomes deciding which ideas deserve generation, which generated clips are usable, which claims need review, and which comparisons are clean enough to trust.

• Hypothesis discipline: each variant should map to a reason for testing, not just a generation variation.
• Naming discipline: the buyer should be able to read results by hook, offer, format, audience intent, and production source.
• Human curation: unusable hands, awkward motion, bad framing, weak product depiction, and off-brand expressions need to be removed before spend starts.
• Provenance checks: teams need to know what inputs, references, voices, likenesses, and claims are being used before the ad enters review.
• Clean comparisons: holdouts or controlled splits matter more when the team can generate many near-variants quickly.
• Finished-ad assembly: captions, voice, pacing, end cards, offer language, and platform-safe exports are still production work.

The media team that benefits from cheaper AI video is not the one that generates the most clips. It is the one that turns cheap raw supply into legible tests.

What should change in an account this quarter

MiniMax’s surge does not justify migrating creative tooling. The crash does not justify abandoning AI video. Neither event tells a buyer whether one model’s output will beat another model’s output in a specific account.

The account-level change is narrower: audit the creative testing system before buying more generation capacity. If the team cannot describe what each video variant is meant to test, more supply will probably create more noise. If the team already has a testing framework but production has been the constraint, low-cost AI video is worth revisiting now.

For small agencies, the immediate upside is not replacing creative judgment. It is making more shots affordable: more hook families, more product-context cuts, more silent b-roll support, more fast refreshes when fatigue appears. That can matter in accounts where the buyer has been rationing video because every new cut required a production request.

For in-house growth teams, the bigger risk is workflow sprawl. One person tests Hailuo clips, another tries a different generator, a designer cleans up exports manually, and the buyer receives assets with no consistent naming or provenance notes. That setup may feel fast in week one and become unreadable by week four.

The ROAS question still has to be earned

Lower generation cost is a creative supply shock. It is not automatic ROAS improvement.

ROAS moves when the finished ad changes user behavior at a cost the account can sustain. That means the creative has to earn attention, communicate the offer, avoid misleading claims, match the landing experience, survive platform review, and beat the existing control or open a new profitable segment. A cheaper clip only helps if it contributes to that chain.

This is where the stock story is most misleading. The surge invited a broad conclusion: AI video is the next ad-tech gold rush. The crash invited the opposite: the hype is over. Neither conclusion is operational enough. The useful read is that the cost of raw creative supply has fallen, while the burden of proof has moved to testing design and post-generation workflow.

A reasonable Monday change is to map the current creative pipeline from idea to spend: who writes the generation brief or script, who approves the concept, who filters the generated clips, who checks rights and provenance, who adds captions and voice, who builds the final variants, and who reads the results. If that map is messy, buying more generation is unlikely to fix it.

Do not change AI video tooling because MiniMax surged or crashed. Reassess the creative testing process because low-cost video generation has shifted the bottleneck from production volume to testing discipline, curation, provenance, and the assembly of raw model output into finished ads.

References

1. MiniMax IPO and 2026 trading timeline.
2. MiniMax valuation and FY2025 financial results.
3. Hailuo 2.3 Fast and Hailuo 02 pricing disclosures.