Amazon in $7B Talks to Acquire Decart: the Real-Time Video Arms Race Begins

$7B talks: Amazon eyes real-time video startup Decart

On October 11 (Sunday), The Wall Street Journal reported, citing people familiar with the matter, that Amazon is in talks to acquire Israeli AI startup Decart for around $7 billion, with a deal possibly closing soon if negotiations hold. Reuters followed up. As of publication, neither Amazon nor Decart has commented — the deal could still fall apart, and that caveat comes first.

The key numbers: in May, Decart raised $300 million (led by Radical Ventures, with Nvidia joining), at a valuation of nearly $4 billion (PitchBook). Five months later, Amazon's reported price is $7 billion — nearly double. More dramatic still: in August, Anthropic also explored acquiring Decart for around $6 billion, then walked away in September after due diligence (Bloomberg). One company, courted by two giants in three months, with the second bid higher — real-time video is moving from "interesting" to "must-have."

Timeline: Decart's 2026 in three leaps

Connect this year's milestones and the $7B talks stop looking sudden:

• May: $300M raise led by Radical Ventures with Nvidia joining, at ~$4B valuation; DOS 2.0 launched alongside a strategic AWS partnership (deep Trainium integration).
• July: Lucy 2.5 shipped, with claimed gains in edit quality, temporal consistency, and physically-aware effects at lower latency.
• August: Anthropic opened acquisition talks at ~$6B (Bloomberg).
• September: Anthropic walked away after diligence.
• October 11: WSJ reports Amazon's ~$7B talks.

Five months, valuation from $4B to $7B. The same curve underneath: real-time video graduating from "tech demo" to "capability cloud vendors must have." Nvidia's May investment, Anthropic's August bid, Amazon's October raise — three giants voting with real money on the same thesis.

What Decart is: a three-piece stack

Founded in 2023 by Dean Leitersdorf, Orian Leitersdorf, and Moshe Shalev, Decart's product line comes in three pieces:

First, Lucy: a real-time video editing model. Its most famous product. Feed in a camera or video stream and Lucy edits it live per your prompt — outfit changes, background swaps, effects — at 30fps with sub-30ms latency. Prompts can be hot-swapped mid-stream without waiting for the next round. The latest is Lucy 2.5 (July 2026), with claimed gains in edit quality, temporal consistency, and physically-aware effects; the realtime version is callable via fal.ai. The matrix also covers VTON (virtual try-on), Restyle (real-time style transfer), and Image (still editing) — each with its own lane.

Second, Oasis: a world model. It generates interactive simulated environments in real time for training and testing robotics and autonomous driving — the "playable AI-generated Minecraft" demo was Oasis. Oasis 3 is on the way — world models and real-time video share the same low-latency inference foundation, which is Decart's clever bit of technology reuse.

Third, DOS (Decart Optimization Stack). It makes the same models run efficiently across Nvidia GPUs, Google TPUs, and Amazon Trainium. This is what separates Decart from pure model shops — it sells the full real-time inference engineering stack, not just weights. For a cloud vendor, that means "buy one Decart, get cross-chip real-time inference," not "buy a video demo."

Figure 1: How Lucy real-time interactive video works. Offline video generation (Sora / Seedance) is "prompt → wait minutes → clip"; Lucy completes every frame within 33ms, with hot-swappable prompts. Source: Decart official docs.

Why Lucy2's tech is hard: three hurdles of real-time generation

Understand these three hurdles and you understand how much of the $7B is technology premium.

Hurdle one: latency. Offline video generation can take its time — ten minutes for a one-minute clip bothers nobody. Real-time interaction can't: 30fps means a 33ms budget per frame; miss it once and the picture stutters. Lucy keeps end-to-end latency under 30ms through co-optimized model architecture and inference stack (the reason the DOS stack exists), not by brute-forcing compute. More compute buys throughput; it doesn't buy latency down — different problems.

Hurdle two: temporal consistency. Frame-by-frame generation fails most often at the seams: shirt colors change, backgrounds jitter, faces drift. Decart's key technique is Self-Anchoring: the model's own previous frame becomes the reference frame for the next, pinning the timeline in place. Third-party tests say Lucy 2.5 stays stable past 60 seconds of continuous generation, where batch-style rivals (including Sora 3) start visually drifting. Sixty seconds happens to be the typical single-session length for livestreams, try-ons, and ad insertions — the metric and the commercial scenario line up.

Hurdle three: chip utilization. The most underrated one. The industry's open secret: most inference workloads run at embarrassingly low utilization, with silicon idling. Lucy2's 80%+ MFU on Trainium3 means Decart's cost per task is structurally lower than rivals'. At $0.02/sec pricing, cost structure is margin — a technology lead eventually becomes ammunition in a price war. What Amazon likely wants is exactly this: the craft of making cheap chips deliver expensive-chip results.

Put together, the answer is clear: Decart's moat isn't one model, it's an engineering system where "low latency + long consistency + high utilization" hold simultaneously. Plenty of teams crack one; few hold all three.

What Amazon is really buying: the Trainium ecosystem + a real-time inference stack

$7 billion for a 2023-founded startup — Amazon is clearly not just buying a video model. Three layers of logic:

Layer one: a showroom for Trainium. Decart is among the first companies to deeply optimize for Trainium, with Lucy2 running on Trainium3. Per Nafea Bshara, VP of Amazon's Annapurna Labs, Lucy2 exceeds 80% Model FLOPS Utilization — "more of the chip's raw power doing real, productive work." In an industry that wastes a lot of silicon, that's a hard engineering metric. Amazon has been pushing its own chips to reduce Nvidia dependence; Decart is the most credible proof point in the Trainium ecosystem.

Layer two: the real-time inference stack. The DOS stack's value is cross-chip portability — one codebase across Nvidia, TPUs, and Trainium. For AWS, acquiring Decart turns "real-time AI inference" into a standard cloud capability sold to media, commerce, advertising, and gaming customers. Decart's CEO demoed Lucy on Trainium3 at AWS re:Invent 2025 — the partnership groundwork was already laid.

Layer three: a ticket to "interactive video" as a category. For two years the video-generation arms race was about "whose clip looks more cinematic." Decart represents the next battlefield — video shifting from "generated to watch" to "real-time interactive." Live try-ons, real-time ad insertion, cloud gaming, and embodied-AI simulation training all need video generated at interactive speed. Amazon is buying a ticket to a new category.

As for why Anthropic walked away after diligence, public reporting gives no reason, and we won't speculate. Just the facts: $6B talked in August, dropped in September, $7B talked in October. The market's price for Decart rose 17% in three months.

The bill: how Lucy2 is priced

Decart is pay-as-you-go with no subscription (official pricing page updated October 14):

• Realtime models: billed per second of active generation. Lucy 2.5 / 2.1 realtime at 720p costs $0.02/sec, Fast mode $0.04/sec, Restyle 2 $0.01/sec.
• Non-realtime video models: billed per generated second, $0.04/sec — twice the price.
• Image models: per generation, $0.01–$0.02 each.

New accounts get free test credits; volume discounts are available for enterprise. Lucy models are also callable via fal.ai at roughly aligned prices, with the benefit of unified billing alongside other models.

Figure 2: Decart Lucy pricing breakdown. Left: per-model rates; right: typical scenario costs — a 10-second outfit change $0.2, one minute of streaming $1.2, one hour always-on $72. Source: Decart official pricing (updated 2026-10-14).

The chart reveals a key product-design constraint: $0.02/sec means Lucy only fits "triggered" scenarios, never "always-on." A viewer tipping for a 10-second outfit change costs $0.2 — totally fine. But a 24/7 AI filter at $72/hour burns money nobody has. Developers have done the math: the right pattern is "viewer-funded triggers, billed by the second," letting GPU cost follow revenue, not airtime. Vibe coders adding real-time video features should do this math first.

Where China stands: ByteDance building, Alibaba publishing, e-commerce already live

ByteDance: strongest offline, realtime on the way. Seedance 2.0 (February) is "director-grade" offline generation — a 60-second cinematic clip in minutes — but it's batch processing, not real-time interaction. Bloomberg reported in September that ByteDance is developing a real-time spatial video model, personally overseen by Zhang Yiming, built on Seedance, targeting ~20fps with ~0.05s latency, cloud-rendered, for interactive virtual worlds on Pico headsets (livestreaming, short dramas, games). An October launch was floated, but the company hasn't commented — treat it as "in development." It's the closest Chinese analog to Lucy and worth watching.

ByteDance's ambition goes beyond one model. 36Kr reported earlier this year that ByteDance ranked its 2026 AI priorities: world models first, then keeping the video-generation lead, improving code, and Doubao monetization. The world-model effort got the company's largest data budget (eight figures in RMB, reportedly 3–4× competitors'), aiming to ship at least one world model before year-end, benchmarked against Google's Genie. Two tracks run in parallel: vision-language-action (VLA) for embodied AI and robotics, and 3D simulation for entertainment and games — the real-time spatial video model belongs to the latter. Cloud rendering carries its own calculus: compute in the cloud means lighter, cheaper Pico headsets, shifting XR competition from "hardware specs" to "models plus cloud capacity" — ByteDance's home turf.

Alibaba: real-time tech, but in paper form. Its Tongyi Lab published Knot Forcing (arXiv:2512.21734) late last year, tackling the core tension of real-time digital-human animation: high quality usually needs slow compute, while real-time easily breaks faces and flickers. Segment-wise generation plus temporal knots plus lookahead navigation lets virtual characters perform fluently, indefinitely. But it's research, not a product API.

One underrated line: e-commerce virtual streamers. China's virtual livestream hosts are already operating at massive scale — but the tech is "driven digital humans" (animating an image/rig), not Lucy's "generative real-time editing." The two routes are converging: driven wins on cost and control, generative wins on expressive ceiling. Decart's $7B price tag is, in a sense, the market pricing the "generative realtime" route.

A vibe coder's checklist: from zero to a working real-time video demo

If this made you want to build, here's the shortest path:

Step 1: try it free. Register on Decart's site — new accounts get free test credits. Or open the public demo at lucy.decart.ai for zero-cost first contact. Verify that "live frame editing" latency is acceptable on your network first — cross-border link latency and the official sub-30ms figure are two different things.

Step 2: get it working via fal.ai. fal.ai aggregates the Lucy family (e.g. decart/lucy2-vton/realtime) with unified billing and solid docs. Open-source projects wiring up Lucy realtime already exist on GitHub — search "decart lucy-2.5 realtime" for WebRTC integration examples; adapting beats writing from scratch.

Step 3: do the math before you code. $0.02/sec = $1.2/min = $72/hr. Convert your scenario into "seconds per trigger": a 10-second virtual try-on is $0.2, a 30-second stream effect $0.6. If the answer comes out "always-on," stop — either reshape the product (triggered), wait for prices to fall, or switch to driven-avatar approaches.

Step 4: design who pays. Real-time video's cost structure makes it a natural fit for "the user pays": viewer tips triggering effects, per-try-on fees in e-commerce, per-session enterprise billing. Build the payment into the product flow instead of subsidizing it afterward. Get this wrong and the slickest demo still loses money.

My take: cloud vendors aren't buying a model, they're buying a ticket to "realtime"

First, the $7B reflects an infrastructure premium. $4B in May → $7B in October: nearly double in five months. What's appreciating isn't Lucy's demo quality, it's the consensus that real-time video is graduating from toy to infrastructure. Anthropic bidding $6B and Amazon $7B corroborate each other: real-time interactive video is no longer a question of "whether," but "who first."

Second, chip co-design is the real hard currency. Note the detail: Decart was among the first to run on Trainium3, and the 80%+ utilization figure came from Annapurna Labs' own VP. Amazon's calculus is clear — it's buying proof that "top-tier results without Nvidia silicon" is possible, which is worth more to the Trainium ecosystem than Lucy itself.

Third, for vibe coders, real-time video APIs are becoming utilities. Decart's official API, fal.ai aggregation, and a possible future AWS integration are dropping the barrier from "train your own model" to "call an API." But $0.02/sec dictates product shape: triggered, viewer-funded, short high-value sessions first. Figuring out "who pays for each second" matters more than figuring out "whether the tech works."

Fourth, a telling contrast: Anthropic walked, Amazon raised. Same company under diligence, yet Anthropic said no in September while Amazon added a billion in October. No reasons are public, but they were never buying the same thing: Anthropic wanted "model capability" (to complement Claude); Amazon wants "chip ecosystem + cloud capability." The same asset prices differently on different strategic boards — a useful footnote for understanding 2026's AI M&A wave: buyers increasingly aren't acquiring "the model itself," but "the land the model grows on."

Risk note: the deal is still just WSJ-citing-sources reporting; both sides are silent and it could collapse (Anthropic's walk-away last month is the cautionary tale). Regulatory approval is another variable — a $7B tech acquisition won't sail through quietly. But whether or not it closes, a $7B spotlight on "real-time interactive video" is already a fact.

Closing: deal or no deal, the category just got its $7B moment

For vibe coders, three lines to remember: real-time video APIs are becoming utilities (Decart's API, fal.ai aggregation, a possible future AWS integration); $0.02/sec means product shape must be triggered, never always-on (do the math before you code); China's counterparts are still on the way (ByteDance's real-time spatial video model is the one to watch). Next time you want to add "real-time outfit change" or "live effects" to a product, you'll know which API to call, what each second costs, and who pays for it — this article's job is done.