MiniMax Goldman Sachs Call: $1 Billion ARR Target by End of 2026
Alina Collins
Goldman Sachs disclosed that MiniMax management has locked in a $1 billion ARR target by end-2026, with revenue projected to leap from $79 million in 2025 to $2.47 billion by 2028 — underpinned by early signs that China's AI pricing war is cooling.
$1 billion ARR — what backs that confidence?
MiniMax's ARR ramp has clear milestones: $100M by end-2025 → $150M by Feb 2026 → doubled again by April, with further acceleration before M3's June 1 launch.
Goldman projects revenue accordingly: $79M in 2025 → $300M in 2026 → $880M in 2027 → $2.47B in 2028.
This means → if the cadence holds, MiniMax goes from sub-$100M to nearly $2.5B in three years — an exponential curve, not a linear extrapolation.
DeepSeek raised prices — why is that an industry turning point?
DeepSeek V4 launches mid-July with peak-hour pricing at 2× off-peak, blended at roughly $0.35 per million tokens (Pro) or $0.12 (Flash).
Goldman reads this as an early signal: the aggressive pricing since April 2026 — some players running at zero or negative gross margins — is shifting toward rationality.
In plain terms = the price war pushed some companies into losses; now even the industry benchmark is raising prices — a sign the "race to the bottom" phase is ending.
Direct benefit for MiniMax: M3's blended price is $0.22 per million tokens, nearly 40% cheaper than DeepSeek Pro — its price-performance edge sharpens as rivals reprice upward.
How does MiniMax maintain high gross margins?
First lever: GPU utilization above 90%. Peak hours serve developers and knowledge workers; off-peak capacity is redeployed for experiments and data curation — no idle waste.
Second lever: an architecture that uses smaller activated parameters for efficient inference. In plain terms = total parameters doubled, but the portion "switched on" per query shrank — so per-inference cost actually fell.
M3 keeps the same price as its predecessor M2.7, but training and inference upgrades deliver more than 2× cost savings, largely offsetting the cost of doubling total parameters.
H3 video model and domestic chips — where do they stand?
The next-generation video model H3 is expected "within weeks." The key upgrade embeds large-language-model capabilities into a diffusion transformer (DiT — an AI architecture for generating video), improving understanding of human motion and basic physics.
MiniMax is recruiting vertical-domain experts to move into feature-film and series production — a step from short-video tools toward professional content.
On the compute side, inference in China already runs heavily on domestic AI chips (ASICs); overseas, localized inference infrastructure covers 200+ countries and regions with a highly diversified client base.
A 400-person company — how does it compete with tech giants?
The entire company has 400–500 employees, with over 80% in R&D — an almost purely technical organization.
Management defines "organizational agility" as their core moat: after OpenClaw emerged, they rapidly commercialized MaxClaw; MiniMax Code shipped quickly too.
This means → MiniMax's strategy is not to outspend incumbents but to out-pivot them — when a new opportunity appears, a small team redirects faster than a big-tech AI lab.
How does Goldman value the stock?
Goldman maintains a Buy rating with a 12-month target of HK$860, implying 141% upside from the current price of HK$356.80.
Two key checkpoints ahead: whether industry pricing continues its return to rationality, and market reception of the H3 video model launch.
This reflects Goldman's core thesis: MiniMax's growth story depends not only on its own product strength, but on whether the broader industry truly moves past the cash-burn-for-volume phase.
Content is for reference only, not financial advice.