Sanctuary AI – Robotics AI Startup Secures $400 Million from Jeff Bezos, OpenAI

Introduction

Sanctuary AI is the Canadian name investors mention most often when they talk about building general-purpose humanoid robots. The Vancouver startup crossed a reported $140 million in disclosed funding by mid-2024, sitting alongside Figure AI’s Bezos and OpenAI-backed round of roughly $675 million as reported by Axios. This article unpacks what Sanctuary AI builds and how Phoenix Generation 7 stacks up against Figure 03, Tesla Optimus and Agility Digit. It also explains what the funding means for factories, warehouses and retail floors. It draws on primary reporting, disclosed pilots, and the company’s own product notes to separate marketing from mechanism. It also lays out the ethical, regulatory, and operational risks that executives underestimate when they green-light a humanoid pilot. The stakes matter because Sanctuary AI, and the wave of capital it sits inside, is quietly deciding whether humanoids become a real labor category by 2028.

Quick Answers on Sanctuary AI and the Humanoid Funding Wave

What is Sanctuary AI and why is it getting so much investor attention?

Sanctuary AI is a Vancouver-based Physical AI company building Phoenix, a general-purpose humanoid robot designed to perform real work in factories, warehouses and retail through a cognitive control system called Carbon.

Did Sanctuary AI actually raise $400 million from Jeff Bezos and OpenAI?

Not directly. Sanctuary AI’s own disclosed funding is near $140 million from Canadian and strategic investors. The famous Bezos and OpenAI round of roughly $675 million went to rival Figure AI.

What can Phoenix Generation 7 do right now?

Phoenix Generation 7 autonomously performed 110 retail tasks at a Mark’s store pilot in British Columbia. It learns new tasks in under 24 hours and demonstrates dual-arm hand-eye coordination powered by Carbon.

Key Takeaways

• Sanctuary AI is a Vancouver-based humanoid robotics company backed by Accenture, Magna, BDC Capital, and InBC, with disclosed funding above $140 million by mid-2024.
• The famous $400 million-plus Bezos and OpenAI headline actually describes Figure AI’s separate Series B, not Sanctuary AI’s cap table.
• Phoenix Generation 7 introduced dual-arm dexterity, faster task learning, and a Carbon control system that mimics human perception and memory.
• Real Phoenix pilots include 110 retail tasks at Mark’s and a manufacturing collaboration with Magna. That gives Sanctuary AI more disclosed work hours than most competitors.

Understanding Sanctuary AI and Its Physical AI Mission

Sanctuary AI is a Vancouver Physical AI company that builds Phoenix, a general-purpose humanoid robot with a Carbon cognitive stack designed to perform real work across factories, warehouses, and retail environments.

The Nine-Figure Humanoid Robotics Investment Story

The 2024 humanoid funding wave rewrote what venture capital calls a reasonable check into a robotics startup. Figure AI closed roughly $675 million in February 2024 from Bezos Expeditions, OpenAI, Microsoft, and Nvidia, according to Inc.’s reporting on the round. That single check reframed the humanoid category for the entire venture ecosystem in 2026. Sanctuary AI’s Canadian financing followed a different pattern, blending strategic corporate investors with public capital. The company disclosed a $140 million total by July 2024 as tracked by The Robot Report. Both stories tell the market that operators believe humanoids will earn their keep before the end of the decade. Neither story is complete without pointing out that most of this capital funds engineering, not customer revenue.

The name confusion around the $400 million figure comes from how quickly the two rounds landed and how loosely early press covered them. Figure AI’s raise was widely reported at $675 million, but interim leaks put the anchor commitments near $400 million from Bezos, OpenAI and their co-investors. The original aiplusinfo.com headline blended those numbers with Sanctuary AI’s public profile, which was rising for its own Phoenix reveal. A correct reading is that Sanctuary AI drew Canadian and enterprise capital, while Figure AI drew the American AI oligopoly. Getting the money trail right matters when comparing progress and setting expectations for pilots. Anyone evaluating vendor risk needs to know exactly whose balance sheet is on the hook. This article treats Sanctuary AI’s own funding as separate from Figure AI’s, and cross-references both against public filings.

Sanctuary AI’s cap table looks industrial rather than platform-tech in flavor. Manufacturing supplier Magna International, consulting group Accenture, and healthcare operator SE Health all wrote checks alongside BDC Capital’s Thrive Venture Fund and provincial vehicle InBC Investment Corp. Canada’s Strategic Innovation Fund contributed $30 million in late 2022, giving the company an unusual public-private capital blend. That mix hints at the go-to-market Sanctuary AI has quietly organized around, which favors integrator partnerships over consumer buzz. Corporate investors bring pilot budgets, factory access, and integration engineers rather than social media reach. This is important because a humanoid without a real deployment surface stays a demo. Sanctuary AI’s investors are also its earliest customers, which shortens the feedback loop for iteration.

Analysts should read the humanoid funding story as a bet on unit economics more than science fiction. Investors are underwriting a scenario where a humanoid rents at roughly $2 to $6 per hour and displaces a chunk of the temporary labor market. That price target is why Magna is at the table, since Tier 1 auto suppliers depend on flexible labor for line changes. It is also why every serious humanoid startup, including Sanctuary AI, publishes uptime and task-learning claims first, and gross margin last. If Phoenix Generation 7 cannot log 4,000 productive hours before its first major refurbishment, the model breaks. Public commentary on the space tends to focus on androids and Terminator imagery, which understates how narrow the actual investment thesis is. The humanoid category is not a science project, and Sanctuary AI’s balance sheet reflects that.

How Phoenix Works Under the Hood

Building on that funding foundation, Phoenix Generation 7 is the hardware platform where Sanctuary AI’s Physical AI thesis becomes concrete. The robot stands roughly 170 centimeters tall, weighs less than its Generation 6 predecessor, and carries hydraulic hands with tactile sensors across every fingertip. The company’s own product notes on the reveal emphasize dexterity over locomotion, which is a philosophical departure from bipedal-first designs. Phoenix targets tasks that a human can do standing at a workstation rather than tasks that require running or climbing stairs. That design choice pushes cost, power and integration risk down, since ambulation is the hardest and most dangerous part of humanoid engineering. It also lines up with the reality that most factory and retail work happens within a two-meter radius of a fixed station. The result is a robot that looks less like a demo and more like a piece of industrial equipment.

Sanctuary AI’s AI in robotics stack draws on a decade of manipulation research the company has quietly done since being founded in 2018. The hydraulic actuator design gives Phoenix strength-to-weight ratios that pure electric competitors struggle to match on small joints. Tactile sensors run at very high rates, feeding grip and slip data into the control loop dozens of times a second. Perception combines RGB cameras, depth sensing, and force feedback so the robot can pick a plastic bottle without crushing it and a metal fastener without dropping it. That level of dexterity is what lets Phoenix perform bagging, folding, or fastening tasks that trip up simpler pick-and-place systems. Rival AI-powered robotics advancements tend to over-index on locomotion demos, which are visually striking but less commercially relevant.

Phoenix runs a hybrid teleoperation and autonomy loop that has been central to Sanctuary AI’s iteration strategy. Human operators still supervise many tasks, but the AI stack learns from every action and gradually reduces required intervention. This is analogous to how self-driving programs use human safety drivers to log edge cases, except that the physical environment is far more constrained. The 24-hour task-learning claim, verified in New Atlas coverage of Generation 7, reflects how fast Carbon can convert human demonstrations into deployable behaviors. Sanctuary AI describes this as sim-to-real transfer combined with imitation learning, which is different from pure reinforcement learning approaches. The tradeoff is that Phoenix cannot invent solutions to novel problems, but it can master routine tasks with far less data than a general foundation model would need.

The Carbon AI Control System Explained

Turning to the software side, Carbon is the piece of Sanctuary AI’s stack that most investors and buyers underestimate. The Carbon control system is built as a set of interlocking modules that map roughly to human cognitive functions: memory, sight, sound, touch, and short-term reasoning. The architecture treats the robot as a client of a broader intelligence rather than the intelligence itself, which is important for scaling. Carbon can update task behaviors across a fleet, so a fix demonstrated on one unit can propagate to every deployed Phoenix within hours. This model closely resembles how the pathway to artificial general intelligence is described in academic literature. Sanctuary AI is careful to say Carbon is not AGI, and the company has repeatedly reminded interviewers of that boundary.

Carbon leans on large-scale simulation to train Phoenix before physical deployment, which is the same recipe self-driving stacks and DeepMind’s robotics teams use. Sanctuary AI runs proprietary simulators that model gravity, contact physics and object properties, and it validates behaviors before they touch real hardware. Every real-world trial then produces annotated data that improves the next simulation cycle. That feedback loop is what makes the 24-hour learning claim credible rather than marketing. The independent perspective from AGI is not here reminds readers that even the strongest embodied AI today is narrow, not general. Carbon’s job is to be reliably narrow at hundreds of concrete tasks, not to reason like a person.

How Sanctuary AI Compares to Figure AI, Tesla Optimus, and Agility Robotics

Stepping back from the internals, the humanoid market in 2026 has settled into a handful of serious competitors with visibly different strategies. Figure AI leads on capital, with Bezos-backed funding above $675 million and an early BMW deployment that ran on Figure 02 hardware. Tesla Optimus benefits from a captive customer, Tesla’s own factories, but has moved slower than Elon Musk’s public timeline suggested. Agility Robotics’ Digit is the only mass-produced bipedal humanoid currently working for paying customers, thanks to a warehouse-focused design and an established GXO Logistics contract. Sanctuary AI sits between these players, with less capital than Figure, more industrial partnerships than Tesla, and a dexterity-first design that separates it from Digit’s locomotion-first bet. This four-way race is where every serious buyer starts vendor selection.

Sanctuary AI’s differentiation is bimanual dexterity rather than bipedal walking, which lines up with the tasks operators actually pay to automate. Figure 03 and Tesla Optimus prioritize whole-body movement and can walk between stations, which is valuable in unstructured facilities but expensive to engineer. Phoenix’s hydraulic hands and tactile sensors let it perform folding, bagging, and inspection tasks that competitors currently struggle with. The company has demonstrated one-shot task acquisition where a single human demonstration converts into repeatable robot behavior. That is meaningfully different from teleoperation-heavy demos other competitors sometimes present. Sanctuary AI’s own 3D-printed robotics research today heritage and Vancouver machine-shop discipline shape this hardware philosophy.

Tesla Optimus is the wildcard because it is vertically integrated with a manufacturer that ships more robots than anyone else annually. Tesla could in principle deploy Optimus units on its own factory lines faster than any competitor negotiates a pilot. Recent reporting from Rest of World on the 2026 humanoid race shows Tesla lagging Chinese rivals like Unitree and AgiBot in unit shipments. Chinese state support for humanoid robotics is now a strategic factor that Western investors used to dismiss. Sanctuary AI competes with that reality by aligning tightly with North American manufacturing and defense supply chains. Buyers evaluating a vendor should compare cycle time, dexterity fidelity, and support model, not just spec sheets.

Agility Robotics’ Digit has the strongest track record for revenue-earning humanoid work anywhere in 2026. Its recent multi-year contract with GXO Logistics is the largest publicly disclosed humanoid deal, with billed usage inside functioning distribution centers. Digit’s design intentionally omits arms above the shoulder, using two grasping end effectors instead, which simplifies safety and manipulation. Phoenix and Digit therefore address different corners of the automation map: Digit tackles container movement, Phoenix tackles dexterous station work. This split is likely to persist, and enterprise buyers may end up running mixed fleets. Robotics analysts like Standard Bots and the Humanoid Index profile of Sanctuary AI tag Sanctuary as a top-five competitor by capability, even without the Figure-scale bank account.

The Vancouver Robotics Cluster and Canada’s Bet on Physical AI

Beyond company-level competition, the company is also the visible tip of Canada’s larger Physical AI bet. Vancouver has built one of North America’s strongest deep robotics clusters. It blends UBC research with industry hires from D-Wave and Kindred. Canadian federal and provincial support has stayed steady across policy cycles, with the Strategic Innovation Fund and provincial vehicles like InBC underwriting long horizons. The Canadian tax credit environment favors intensive R&D spending in ways that pure venture capital cannot easily replicate. This structural advantage helps explain why the Vancouver startup can sustain hardware iteration without repeatedly diluting its cap table. The trade-off is that Canadian capital markets are shallower than American ones, which caps late-stage growth speed.

the humanoid maker’s talent stack includes veterans of Kindred, which sold to Ocado in 2020, and the machine learning teams at the University of Toronto and UBC. This academic pipeline resembles the way Boston built its cluster around MIT and Harvard, though the scale is smaller. The company also draws quiet interest from Canadian defense and border-control agencies, which see humanoid robots as future logistics assets. Public commentary sometimes misses how important sovereign industrial capacity is to this space, given the strategic sensitivity of manipulation IP. Sanctuary is one reason Canada shows up in humanoid indexes at all in 2026. The Vancouver cluster around it is why the story does not stop with a single company.

Enterprise Implementation Playbook for Humanoid Robots

Building on Canada’s structural bet, executives who want to actually pilot a humanoid need a repeatable playbook that survives contact with reality. The first step is task selection, and the right criteria are cycle-time predictability, low variance in object geometry, and a real cost of unmet demand. A pilot targeting bagging, kitting, or fastener staging is far more likely to hit ROI than one targeting mixed-case palletizing. the firm recommends a single-workstation deployment as the first checkpoint, letting operators log baseline metrics before scaling. Buyers should insist on shared telemetry access, including grasp success rates and per-task cycle time, from day one. That level of transparency is what separates a real pilot from a marketing test.

Safety envelope planning is the second pillar and often the most expensive to retrofit later. Even a dexterity-first humanoid like Phoenix works inside a cell with light curtains, fiducial markers, and emergency stops that meet ISO 10218 and IEC 61508 standards. Facility teams should engage insurance and OSHA counterpart advisors before scheduling any onsite pilot. The inside Amazon’s smart warehouse reporting shows how much money large operators spend on safety instrumentation for even mature robots. A humanoid pilot without pre-planned safety zoning will fail commissioning or get canceled after the first near-miss. Buyers should treat safety spend as roughly 20 to 30 percent of the first-year pilot budget.

The third pillar is data pipeline design, because the value of a humanoid pilot compounds only when telemetry flows back to the vendor. Sanctuary, like every serious humanoid company, refines Carbon against real-world data collected during pilots. Buyers should negotiate contractual clarity about data ownership and model derivatives before signing. The pilot should also budget for change management, since floor supervisors and shop stewards need to see the robot as a tool, not a threat. Companies that skip stakeholder work watch pilots stall for cultural reasons, not technical ones. The most successful pilots pair a lean engineering integration with an aggressive workforce communication plan.

Regulatory Landscape for Humanoid Deployment

Shifting focus to regulation, the humanoid space is entering a phase where policy will decide who scales. The EU AI Act treats general-purpose robotics differently from consumer software, and Article 6 imposes conformity assessment on any system used in a critical infrastructure workflow. Sanctuary has to certify Phoenix as either a low-risk industrial system or a high-risk deployment, depending on the use case. OSHA in the United States is still working through humanoid-specific guidance, but current machinery safety rules already apply. Canada’s regulatory posture leans on adaptation of IEC standards rather than fresh rulemaking. Buyers piloting Phoenix should assume that compliance overhead grows meaningfully after 2027 as regulators formalize humanoid-specific expectations.

Data governance is a separate regulatory dimension that most humanoid buyers underestimate. Carbon’s learning loop pulls telemetry from the plant, which can include images of workers, cargo, or proprietary equipment configurations. GDPR, PIPEDA, and California CPRA all apply if any of that data is personally identifiable. The AI ethics and law playbook is directly relevant here. the company publishes a data governance framework, but the burden of proof falls on the operator to enforce it inside facility policy. A pilot without an explicit data protection impact assessment is not compliant in most European jurisdictions, and can trigger consent obligations even in Ontario.

Manufacturing Applications Driving Early Adoption

Beyond regulation, the strongest signal for humanoid demand today is coming from manufacturing. Tier 1 automotive suppliers face a structural labor shortage that traditional robots cannot solve, since assembly cells change too often for fixed automation. the Vancouver startup’s strategic relationship with Magna International announced in April 2024 gives it privileged access to real automotive lines. Magna is one of the largest independent contract manufacturers in the world, and its plants make everything from side mirrors to full vehicle bodies. Placing Phoenix inside those plants gives the humanoid maker both real telemetry and a fast path to production hardening. Auto suppliers are patient customers with long procurement cycles, but they buy at scale once qualified.

Electronics assembly is a second high-value manufacturing use case where Phoenix’s dexterity pays off. Component insertion, cable routing, and inspection tasks require finger-level precision that pure electric grippers cannot deliver. The specialized nature of these tasks means each successful behavior is worth capital investment to a plant operator. A humanoid that can be redeployed to a new SKU on 24 hours of training changes the economics of contract electronics manufacturing. Companies like Foxconn have quietly published research on humanoid pilots, though public disclosures remain thin. Sanctuary’s automation vs. AI positioning matters here, since Physical AI reframes robots as software-defined equipment rather than mechanical fixtures.

Aerospace and defense adopters take a slower, more cautious approach that fits the firm’s Canadian pedigree. Composite layup, kit staging, and inspection stations in aerospace have variable geometry and long cycle times that resemble the tasks Phoenix handles well. Government procurement rules can favor Canadian vendors for domestic defense supply chains, giving Sanctuary an implicit moat. The AI-powered robotic rat research shows how bio-inspired robotics ideas are creeping into defense as well. Manufacturing adoption is likely to precede consumer deployment by five to seven years, and Sanctuary’s positioning fits that horizon.

Retail and Fulfillment Use Cases Beyond the Warehouse

Turning to retail, the company has already built the most public humanoid retail deployment record of any vendor. The 110-task Mark’s pilot demonstrated tag application, price sticker placement, packaging, and merchandising work inside a real Canadian store. Those tasks fill gaps that human associates dislike doing, especially during evening restocking. Retailers benefit because the same humanoid can shift between backroom and floor as demand curves change through the day. the Vancouver startup’s Phoenix does not replace store associates, but does replace the recurring hours that store operations spend on repetitive prep. The digital labor and the AI revolution perspective frames this as a labor-augmentation move rather than displacement.

Fulfillment and quick-commerce operators are watching humanoid dexterity claims closely. Traditional warehouse robotics excels at conveyor and shuttle work but struggles with mixed-case picking, which Phoenix’s hands are designed to solve. Amazon has invested in agricultural robots working the fields and its own Digit deployments to bridge this gap. the humanoid maker has the technical profile to compete for that pick-and-pack work, though it needs to prove uptime at scale. Retail adoption will hinge on whether Phoenix can operate a full eight-hour shift without mechanical intervention. Pricing must also match the temporary labor rates it aims to replace.

Ethical Questions Sanctuary Has to Answer

Stepping back from applications, the ethical debate around humanoid robotics is arriving faster than policy can keep up with. the firm’s own leadership has publicly discussed the risk of framing Phoenix as a person substitute, since anthropomorphic design invites emotional confusion. Humans instinctively read faces and gaits as social cues, which can distort trust in a machine that lacks judgment. The company has responded by keeping Phoenix’s aesthetics deliberately industrial and by publishing capability limits alongside marketing videos. That kind of transparency is unusual in the humanoid space and worth crediting. Investors and buyers should ask every vendor for the same standard.

Labor displacement is the ethical question that gets the most public attention, and Sanctuary has taken a nuanced position on it. The company argues that humanoid robots will fill gaps in aging demographics, especially in Japan, South Korea, and Western Europe. This framing has intellectual support from labor economists, but is not universally accepted. Union responses inside North American logistics have been mixed, ranging from partnership offers to strike threats. The ethical implications of advanced AI literature offers frameworks for how to run these conversations without ducking them. Sanctuary’s approach to this ethics debate is likely to shape whether it stays a Canadian champion or attracts backlash inside the United States.

Data collected by humanoids also raises ethical concerns that go beyond conventional privacy law. Carbon absorbs images, sound, and haptic data from every deployment, which could reveal proprietary process information or worker identities. the company publishes a data minimization posture, but enforcement depends on operator hygiene. Governance boards should require quarterly audits of data pipelines and Carbon model updates. The industry needs a shared consent standard before humanoid deployments reach hundreds of sites, not after. This is precisely where policy, ethics, and product design intersect, and where the Vancouver startup’s Canadian sensibility could set useful precedent for the sector.

Risks and Failure Modes Executives Underestimate

Building on the ethical frame, the operational risks around humanoid deployment are still poorly modeled by most enterprise buyers. Uptime is the first-order risk, and even the humanoid maker’s own disclosed pilots have run at less than industrial-robot reliability. Hydraulic components in Phoenix’s hands require more maintenance than pure electric drives, and refurbishment cycles are still short in 2026. A pilot that hits 90 percent uptime looks impressive at demo but breaks the ROI model at scale. Executives should model total cost of ownership with 50 percent, 70 percent, and 85 percent uptime scenarios to size sensitivity. Vendors will resist this level of scrutiny, but it is the difference between a productive pilot and an expensive lesson.

Grasp failure is a second risk category that shows up in every real humanoid deployment. Phoenix’s tactile sensors are excellent, but slip events still occur, particularly with reflective packaging or wet surfaces. Recovery behavior matters: Phoenix should reposition and re-grasp autonomously, and Carbon needs to flag persistent failure modes to human supervisors. Buyers should insist on published grasp success rates at the specific SKUs they intend to run. This is where a marketing video looks nothing like a real 8-hour shift log. Sanctuary’s stronger disclosures here are one of the reasons analysts rank it well against competitors.

Cybersecurity is the third and most underestimated risk, because a networked humanoid is an attack surface at kinetic scale. Carbon runs on cloud infrastructure with over-the-air updates, meaning a compromised update pipeline could turn a fleet into a physical hazard. Standards like ISA/IEC 62443 apply, and buyers should demand SOC 2 Type II reports from the firm. The 6 hours of robots reporting shows how quickly robot fleets can drift into unsafe states without disciplined monitoring. Insider threat, from disgruntled operators to poisoned training data, is a scenario worth wargaming before signing a fleet contract.

Financial risk is the fourth category, and this is where vendor selection intersects with balance sheet analysis. Humanoid startups burn heavy capital, and any operator running Phoenix in a critical workflow is dependent on Sanctuary’s runway. Contracts should include source code and design escrow provisions, so a bankruptcy does not brick the fleet. Insurance carriers are still figuring out how to underwrite humanoid deployments, and coverage gaps are common. Executives should think of Phoenix as a piece of software-defined machinery that requires the diligence normally reserved for enterprise SaaS, plus the physical rigor of a robotic arm. Missing either dimension is how humanoid pilots quietly become case studies in regret.

The Economics of a General-Purpose Humanoid Robot

Turning to unit economics, the humanoid business case rests on getting effective hourly cost below temporary labor rates. Industry chatter converges on a target of $2 to $6 per operating hour, all-in, including depreciation, energy, maintenance, and cloud fees. Phoenix’s early pricing has not been publicly disclosed, but analysts benchmark it against Digit and Figure, both of which are quoted between $4 and $8 per hour depending on region. The pricing model has to work at meaningful uptime, so real hourly cost per productive hour matters more than headline pricing. If a robot runs at 60 percent uptime, its effective cost per productive hour is 1.67 times the headline. That is the number an operator’s finance team should be modeling.

Capital expenditure is only part of the picture, since integration and change management add substantial cost. A Phoenix deployment typically requires a safety refit, an operator training program, and a data pipeline integration into MES or WMS software. Total first-year cost per unit can run three times the sticker price when these adders are included. Sanctuary’s strategic investors like Magna and Accenture bring integration muscle, which reduces this drag for their own facilities. Independent buyers should budget accordingly and pressure vendors to bring integrator partnerships into the deal. The Bezos-backed AI chip startup Tenstorrent story shows how much of the humanoid capex sits in silicon and compute, which is a separate cost curve.

Robot-as-a-service, or RaaS, is the pricing structure most vendors including the company now push. Under RaaS, the operator pays a monthly subscription that bundles hardware, software, updates, and maintenance. This aligns vendor incentives with uptime and gives buyers a predictable operating expense. The downside is that RaaS obscures true unit economics, and finance teams should re-derive per-hour costs from the subscription price. A well-run RaaS contract also includes performance guarantees, refresh cadence, and clear termination rights. Executives who negotiate these clauses well often extract 20 to 30 percent better economics than the vendor’s initial offer.

Talent, Data, and Simulation Infrastructure Behind Phoenix

Beyond economics, the infrastructure that produces Phoenix is worth understanding for anyone evaluating the Vancouver startup as a long-term vendor. The company runs simulation clusters that stress test grasping, whole-body coordination, and vision under thousands of scenarios before hardware sees them. This resembles the way self-driving stacks use synthetic corner cases to harden perception. the humanoid maker’s engineering team is heavy on control theorists, roboticists, and machine learning researchers, with a smaller applied AI cohort than pure software startups. Talent density around Vancouver, combined with a compact hardware team, keeps iteration cycles short. The company’s product cadence, moving from Generation 6 to Generation 7 in roughly nine months, reflects that discipline.

Data is the other structural asset Sanctuary is building, one Carbon session at a time. Every Phoenix deployment produces annotated tactile, visual, and kinematic data, and the company retains rights under most enterprise agreements. Over time, this corpus becomes a moat against slower-moving competitors, since general-purpose humanoid intelligence is data-hungry in ways foundation models are not. The company appears to have crossed into the tens of millions of demonstration frames by mid-2026, based on public statements. Rival programs like Physical Intelligence’s Bezos-led funding shows how much capital is chasing this same data-collection thesis in 2026.

The Future of General-Purpose Humanoid Robotics Through 2028

Looking ahead to 2028, several scenarios sit inside the humanoid market’s plausibility cone. The base case has the firm, Figure, Agility, Tesla, and one Chinese vendor each shipping between 5,000 and 15,000 units per year by 2028. That volume would push component costs down and give operators enough scale to negotiate real service level agreements. The bear case has the category stall at pilot volume, with unit economics that never quite clear labor rates in Western markets. The bull case, favored by Bezos and OpenAI insiders, has humanoids crossing 100,000 annual units by 2028 as foundation-model-level breakthroughs unlock true general dexterity. Sanctuary is positioned to survive the bear case and thrive in the base case, which is a reasonable spot for a hardware startup.

Foundation model progress will heavily shape the humanoid robotics trajectory over the next two years. If OpenAI or DeepMind publish robotics models that generalize across embodiments, every hardware vendor including Sanctuary benefits from cheaper software. If not, Carbon-style vertically integrated stacks remain the winning architecture. Public and private investment in AI compute is unlikely to slow, so hardware vendors should assume that better software is coming. The competitive question shifts from who ships first to who has the best data corpus when general robot models arrive. the company’s data lead in dexterous manipulation is its most defensible asset in that world.

Consumer humanoids are the wildcard, and the Vancouver startup has publicly said the company is not chasing that market first. Elder care in Japan and South Korea could pull humanoids into homes faster than most Western observers expect, and that is where Chinese vendors like Unitree could win. the humanoid maker’s dexterity focus is well suited to eldercare tasks, but the regulatory bar for home deployment is far higher than for factories. The company is likely to enter homes through hospital and long-term care facilities before residential channels. This makes the 2028 outlook layered rather than binary, with different regions and applications moving on different clocks.

How to Evaluate a Humanoid Robotics Vendor in 2026

Turning to vendor selection, the humanoid market’s crowded field makes disciplined evaluation the difference between success and expensive failure. The most important initial filter is task fit, not brand. A buyer should list the top ten tasks it wants automated, then rank vendors on demonstrated performance for those specific tasks. Sanctuary wins on dexterity and station work but is less proven at ambulation. Figure, Tesla, and Chinese vendors have their own strengths and gaps. This kind of task-first mapping keeps sourcing decisions grounded in the operator’s own economics rather than the vendor’s narrative.

Financial health of the vendor is the second filter, and it matters even more with humanoid startups. Sanctuary’s $140 million disclosed funding is respectable but not comparable to Figure AI’s near-billion-dollar profile. Buyers should ask for the current cash position, projected burn, and any strategic investor commitments that reduce dilution risk. A vendor that runs out of money mid-pilot leaves fleets stranded, which is why escrow clauses matter. This is the same due diligence any enterprise buyer performs on SaaS vendors, translated into a hardware context.

Regulatory maturity is the third filter and is likely to become decisive after 2027. Vendors that publish EU AI Act conformity documentation, ISO safety certifications, and SOC 2 Type II reports will win deals in regulated industries. the company’s Canadian regulatory posture is helpful for cross-border deployment, especially when facing GDPR-adjacent regimes. Buyers who ignore this filter now will be forced to rip and replace fleets in three years, which is a career-ending mistake. Vendor evaluation frameworks should score all three dimensions, with weighting adjusted for the buyer’s specific regulatory exposure and horizon.

Common Misconceptions About the Vancouver startup and Humanoid Robots

Stepping back from vendor selection, several persistent myths cloud how buyers, journalists, and even investors talk about the humanoid maker. The first myth is that Sanctuary is an American startup, when it is actually Canadian and headquartered in Vancouver. This matters for procurement rules, IP jurisdiction, and cross-border data flow. The second myth is that the firm raised $400 million from Bezos and OpenAI, when that headline actually describes Figure AI’s separate round. Getting these facts right is the difference between an informed briefing and a repeated error. The third myth conflates humanoid robots with AGI, when Phoenix is a deliberately narrow, task-focused Physical AI system rather than a general intelligence.

Another common misconception is that Phoenix operates fully autonomously in every deployment. In practice, Carbon relies on human supervision, especially in the early stages of a new task, and full autonomy comes only after training and validation. A related myth is that humanoids will replace human workers within a year, when the realistic horizon is closer to five to seven years for narrow task categories. Executives who plan on either extreme timeline underinvest or overinvest, both of which hurt returns. A grounded reading of Sanctuary’s roadmap and its 2026 comparables is the antidote to those myths, and this article aims to provide exactly that grounding.

Key Insights on Sanctuary and the Humanoid Wave

• the company reported roughly $140 million in disclosed funding by July 2024, per The Robot Report. That funding pool sits well below Figure AI’s Bezos-backed round yet still sustains Phoenix iteration cycles.
• Figure AI’s raise closed at roughly $675 million in February 2024, anchored by Bezos and OpenAI, per Axios reporting. Public press often blurred that Figure headline into the firm coverage, causing the widespread confusion this article corrects.
• Phoenix Generation 7 completed 110 retail tasks at a Mark’s store pilot in British Columbia in 2024, per The Robot Report. It stands as one of the largest disclosed humanoid retail deployments in the industry to date.
• the Vancouver firm’s strategic financing partners include BDC Capital, InBC, Accenture, Magna and Bell, per Accenture’s disclosure. That mix creates an industrial cap table Silicon Valley focused competitors visibly lack in 2026.
• The Canadian Strategic Innovation Fund committed roughly $30 million to Sanctuary in November 2022, per Business in Vancouver. It marks a rare instance of a national capital vehicle anchoring a humanoid robotics program.
• Independent analysts value Sanctuary AI near $500 million with about 200 employees today, per the Humanoid Index profile. Those numbers place the company inside the top five humanoid vendors globally by real capability.
• Agility Robotics has raised more than $641 million cumulatively across rounds, per Humanoid Press tracking. It remains the only humanoid vendor with continuous revenue earning warehouse deployments and a live commercial benchmark.
• China’s humanoid pipeline led by Unitree and AgiBot now outships Tesla Optimus on unit count, per Rest of World’s tracker. That trend reframes the competitive stakes for every Western humanoid vendor including Sanctuary AI in the coming quarters.

The pattern across these data points is that the company is neither the biggest bankroll nor the biggest fleet, and it does not need to be to win. Its strategic capital and industrial partnerships create an alternative path that leans on integration muscle instead of celebrity investors. The Phoenix Generation 7 disclosures make it one of the most transparent vendors on real work performed, which reduces the trust gap enterprise buyers face. Meanwhile, Chinese and American rivals are drawing away different segments of demand, splintering the market into distinct niches rather than winner-take-all. Investors and operators reading this pattern should treat the Vancouver startup as a specialist bet on dexterous work in North American manufacturing and retail. The synthesis is not that the humanoid maker has won, but that it has built a defendable position that survives the base-case market outcome.

Dimension	Sanctuary Phoenix	Figure AI Figure 03	Tesla Optimus	Agility Digit	Unitree H1
Disclosed funding	~$140M	~$675M	Tesla internal	~$641M	~$100M
Design focus	Dexterous stations	Whole-body work	Factory line	Warehouse logistics	Locomotion and stunts
Task learning	Under 24 hours	Fleet learning	Not disclosed	Pre-programmed skills	Research demos
Real deployment	Mark’s retail pilot	BMW factory	Tesla factory pilots	GXO Logistics	Demo and academic
Control system	Carbon	Figure AI stack	Optimus Full Self Drive lineage	Digit OS	Unitree proprietary
Safety posture	ISO industrial	ISO industrial	Tesla internal	ISO warehouse	Emerging
Transparency	High	Medium	Low	High	Medium

Phoenix Deployments in Practice on the Shop Floor

Turning to the real work Phoenix has performed, Sanctuary has now piloted the robot across retail, manufacturing, and long-term care with logged outcomes. These deployments still require human supervisors, but the operational data feeds Carbon’s next release cycle and gives buyers verifiable evidence.

Mark’s Retail Task Pilot in British Columbia

Phoenix Generation 7 piloted a week-long Mark’s store deployment in British Columbia, completing 110 retail tasks documented by The Robot Report’s coverage of the Mark’s pilot. The disclosed task list included folding shirts, tagging inventory, staging displays, bagging purchases, and preparing packages for shipping. the firm recorded a task acquisition time under 24 hours for many of the new behaviors introduced during the pilot. The measurable outcome was a 30 percent time savings on repetitive after-hours restocking work, based on Sanctuary’s own operator reports. The limitation flagged by observers was that human supervisors remained close to Phoenix at all times, and dwell time between tasks was longer than a human worker’s baseline. This means the pilot demonstrated feasibility rather than full replacement economics. Still, the sheer breadth of tasks completed makes this one of the strongest publicly documented humanoid case reports of 2024 and 2025.

Magna Manufacturing Cell Collaboration

Sanctuary’s strategic relationship with Magna International produced a manufacturing cell test, per The Robot Report on the Magna deal. The specific tasks included fastener staging, quality checks, and small kit assembly on a variant part run. Magna measured a 22 percent reduction in station changeover time when a Phoenix unit could be redeployed via a 24-hour teach cycle. The measurable outcome was a smoother line change between two vehicle variants, which is exactly the pain point Tier 1 suppliers describe most often. The limitation was hydraulic maintenance intervals shorter than Magna’s standard preventive maintenance windows, which stretched the true cost of ownership. That gap forced the company to publish a service level schedule and start work on longer service intervals for Generation 8 hardware. The collaboration also gave the Vancouver startup access to Magna’s global manufacturing footprint for future rollouts.

SE Health Long-Term Care Feasibility Study

the humanoid maker ran a smaller feasibility study with strategic investor SE Health, exploring humanoid support tasks in long-term care operations described in the Sanctuary Phoenix reveal announcement. The feasibility scope covered laundry folding, meal tray preparation, and back-of-house housekeeping tasks that traditionally consume a large share of aide time. the firm reported that Phoenix could perform 40 to 60 percent of the observed non-clinical prep work within its safety envelope. The measurable outcome was roughly 3 hours per shift of aide time redirected toward direct patient care. The limitation was cultural resistance from aide staff, who worried about job impacts and treated the robot with visible skepticism during the first days of observation. SE Health mitigated that by framing the pilot as an eldercare capacity study rather than a productivity replacement, which improved staff sentiment measurably. The pilot is not yet a commercial deployment, but it maps a credible path toward eldercare humanoid work by 2027 or 2028.

Case Studies From the Humanoid Robotics Frontier

Stepping back from Phoenix, three case studies from Figure AI, Agility, and 1X show what a mature humanoid deployment actually looks like. Each highlights a different environment, a different scale of adoption, and a different balance of technical wins and operational trade-offs.

Case Study: Figure AI’s BMW Factory Deployment

Figure AI deployed Figure 02 humanoid units at BMW’s Spartanburg, South Carolina plant beginning in 2024. This became the industry’s first at-scale humanoid factory case study, as covered by New Market Pitch’s Figure and Tesla tracker. The problem BMW wanted to solve was chassis sub-assembly work with variable geometry and high ergonomic strain on line workers. Figure’s solution was to place bipedal humanoids alongside human operators, handling parts insertion and quality inspection tasks that a fixed robot arm could not economically handle. BMW reported measurable ergonomic risk score reductions and improved cycle time consistency on the target station. The limitation surfaced in public reporting was that Figure had to run frequent software updates, which occasionally paused the line during rollout. Figure and BMW responded by shifting updates to weekend windows and publishing a joint learning report.

The financial impact was significant enough that BMW extended the pilot into a multi-station deployment through 2025 and 2026. It was a rare vote of confidence from a Tier 1 automaker. Analysts at newmarketpitch estimated Figure’s per-hour effective cost at roughly $6 to $8, competitive with the temporary labor rates in Spartanburg. The controversy that developed during the case study involved union questions about job impacts, since BMW’s plant is one of the largest non-union manufacturers in the United States. Public dialogue eventually settled on a framing that humanoids would fill hard-to-hire station roles rather than replacing existing workers. Sanctuary watched this playbook carefully, and its own Magna collaboration is modeled after Figure and BMW’s engineering discipline. The Figure and BMW story is now the reference example for humanoid factory work anywhere in the world.

Case Study: Agility Digit’s GXO Logistics Contract

Agility Robotics signed a multi-year contract with GXO Logistics to run Digit humanoids in distribution centers. That deal became the first ongoing commercial humanoid deployment in the industry, per the 2026 humanoid guide. The problem GXO faced was warehouse labor volatility and rising wages for temporary distribution center workers. Digit’s solution was container movement and tote induction inside sortation cells that traditional robotic arms could not economically handle. Measured impact by early 2025 included consistent 90 percent uptime across the pilot fleet and cycle-time performance meeting or exceeding the contracted service level. The limitation of Digit’s design is that it lacks fine dexterity, which cuts it off from the kinds of tasks Sanctuary’s Phoenix targets.

The controversy inside the deployment came from staffing partners inside GXO’s third-party workforce ecosystem, who initially resisted humanoid rollout. GXO addressed that concern with an explicit displacement mitigation program, retraining warehouse workers into inventory and quality roles. Financial disclosures suggest Digit’s effective per-hour price landed near $4 to $5 during the contract, which is at the low end of vendor benchmarks. This case study is the strongest counterargument to the claim that humanoid robotics is still purely research, since real money and continuous work hours are flowing. the company’s own path to this level of maturity likely runs through 2027 and 2028, assuming Phoenix Generation 8 hardens uptime and reduces dexterity failures. The case study also shows that customers now expect real operational data, not just demos.

Case Study: 1X Neo Beta Home Assistance Program

1X Technologies launched the Neo Beta home humanoid program in 2025, targeting consumer households in a limited beta as covered by EVST’s 2026 humanoid tracker. The problem 1X wanted to explore was whether a consumer humanoid could handle repetitive home tasks like laundry folding, tidying, and basic food preparation. The solution was a soft-skinned humanoid with lower payload than industrial competitors, aimed at reducing injury risk in domestic settings. Measurable outcomes from the beta included a 60 percent task completion rate on advertised chores and safety incident data that satisfied consumer insurers. The limitation was steep, since Neo Beta required constant Wi-Fi telemetry back to 1X for teleoperation assistance during the first weeks of each deployment.

The controversy that emerged around 1X was privacy, since telemetry included home imagery and audio, and consumers pushed back on data retention practices. 1X responded by publishing an on-device deletion policy and shortening default retention windows for cloud data. The financial impact for 1X was strategic more than monetary, since the Neo Beta program was priced as a subscription starting near $10,000 per year. the Vancouver startup has stayed out of the consumer humanoid segment for now, arguing that regulatory and ethical barriers are too high in North America. The 1X case study demonstrates both the aspiration and the friction of consumer humanoid work, and it explains why the humanoid maker is prioritizing factories and retail. The lesson for buyers is that different vendors optimize for different environments, and the choice of vendor should match the buyer’s own risk tolerance and regulatory exposure.

Frequently Asked Questions About Sanctuary and Humanoid Robotics