The first two posts in this series covered software AI employees: Part 1 by industry and role, Part 2 the small-business stacks. This post is the missing leg of the stool. Software AI employees handle the digital work. Humanoid robots — built on the same foundation-model substrate — handle the physical work. And the rate at which the two converge is what every operations-heavy industry needs to plan around for the next four years.
This isn't speculation. Banks have published forecasts (Goldman Sachs, Bank of America, Morgan Stanley). Hardware companies have shipped commercial deployments (Figure at BMW, Agility's Digit at Amazon and GXO). Foundation-model labs (NVIDIA, Physical Intelligence, Skild AI) have raised billions to build the brain layer. The question for an operator isn't whether this is happening — the question is what to plan for in 2027, 2028, 2029, and 2030, and why those specific years.
The framing for this post: I'll give the cited current state, then walk year by year through what published roadmaps and capital flows imply. Every claim is footnoted. Forecast dates beyond 2026 are vendor or analyst projections — labeled clearly — not certainties.
The convergence thesis in one paragraph
A 2018 robot ran on hand-crafted control code and a custom perception pipeline; each new task required a new project. A 2026 robot — Figure 02, Apptronik Apollo, Agility Digit, Tesla Optimus Gen 2 — runs on a vision-language-action (VLA) foundation model that gets better when the model gets better, the same way a software AI employee improves when the underlying LLM does. NVIDIA's Project GR00T (announced March 2024) is the explicit foundation-model layer for humanoid robots.[^1] Physical Intelligence ("PI") raised $400M at a $2.4B valuation in November 2024 to build a single robot foundation model.[^2] Skild AI raised $300M in mid-2024 from Jeff Bezos, Lightspeed, and others for a generalist robot brain.[^3] What this means: software AI and physical-robot AI now improve on a shared rate-of-progress curve. The 2027 humanoid is not the 2026 humanoid plus 12 months of mechanical engineering — it's the 2026 humanoid plus 12 months of model improvement, which has been moving roughly 4× to 10× faster than mechanical engineering historically does.
State of the field — what's actually deployed in mid-2026
A short, cited inventory.
- Figure AI raised $675M at a $2.6B valuation in February 2024 (investors include Microsoft, OpenAI, NVIDIA, Jeff Bezos).[^4] Their Figure 02 humanoid is in pilot deployment at BMW's Spartanburg, South Carolina plant — the partnership with BMW was announced January 2024 and Figure has publicly demonstrated unsupervised task execution on the line.[^5]
- Agility Robotics' Digit has been operational at Amazon warehouses (since 2023) and GXO Logistics facilities (since 2024) for tote handling and material movement. Agility's CEO publicly stated they shipped commercial units to GXO in 2024 under a multi-year contract.[^6]
- Apptronik's Apollo humanoid has commercial pilots with Mercedes-Benz (announced March 2024) and GXO (announced October 2024 for a multi-year deployment).[^7]
- Tesla Optimus Gen 2 was demonstrated December 2023; Elon Musk publicly stated production aspirations targeting 2026, with internal Tesla pilot deployment for materials handling on Tesla's own production lines.[^8]
- Boston Dynamics' all-electric Atlas was unveiled April 2024, replacing the hydraulic Atlas. Pilot programs at Hyundai (Boston Dynamics' parent) and other industrial customers are public.[^9]
- Boston Dynamics' Spot has thousands of units deployed across construction, oil-and-gas, and utility customers (BP, Hensel Phelps, etc.).
- 1X Technologies (Norway/US humanoid; OpenAI co-investor) raised $100M Series B in January 2024 and has demonstrated their NEO Beta home/light-commercial humanoid.[^10]
- Unitree (China) ships the H1 humanoid commercially at ~$16,000 per unit (publicly listed); the lower-cost Go2 quadruped sells under $3,000 — the first time a price-competitive consumer-tier robot has shipped at volume.[^11]
The brain layer:
- NVIDIA Project GR00T: foundation model + Isaac robotics platform for training humanoid skills via simulation.[^1]
- Physical Intelligence (PI): robot-agnostic foundation model; their "π0" model (April 2024) demonstrated novel skills (laundry folding, table bussing) zero-shot on previously-unseen hardware.[^12]
- Skild AI: generalist robot brain, $300M raised mid-2024.[^3]
- Google DeepMind Gemini Robotics announcements (2024-2025): vision-language-action models for general-purpose robots.[^13]
The forecast: years one through four, with citations
The framework I'll use: each year section has a cited driver (what's already in motion), a vendor / analyst projection (what's been publicly stated), and a practical SMB-or-individual takeaway.
Year 1 (calendar 2027): pilot-to-production transition
Cited driver: Figure AI publicly stated 2025–2026 as the commercialization window for Figure 02 broadly across BMW and additional automotive customers; Apptronik's Mercedes pilot is on a multi-year contract maturing into wider deployment; Agility's GXO contract scales unit count year-over-year per their public statements.[^5][^6][^7]
Analyst projection: Goldman Sachs Research updated their humanoid robot market forecast in January 2024, projecting the TAM at $38 billion by 2035 with 1.4 million units shipped annually at maturity. They explicitly modeled that 2025–2027 is the "early adoption" phase with thousands-not-millions of units in deployment.[^14] Morgan Stanley projected (April 2024) that humanoid robots could ship >250,000 units by 2030 as the inflection point.[^15]
Practical takeaway: In 2027 you'll see the first humanoid-robot deployments in mid-tier logistics warehouses, not just BMW-grade automotive plants. Cost per humanoid in the $30K–$80K USD range becomes plausible (down from current $50K–$150K cost basis at low volume). For an SMB owner: budget for it, don't deploy yet — the economic case sharpens in 2028.
Year 2 (calendar 2028): regulatory and price tipping
Cited driver: A 2-generation cycle on humanoids (Figure 02 → Figure 03 → Figure 04) compounds the model-driven capability gain. NVIDIA's GR00T roadmap targets a generally-available robot foundation model, and Physical Intelligence's π-series is expected to release multiple successor models.[^1][^12] OSHA, the EU Machinery Regulation (revised 2023, in force 2027), and equivalent regulators are explicitly scoping AI/robot safety standards on a 2027-2028 timeline.[^16]
Analyst projection: Bank of America Global Research published in May 2024 a forecast that the humanoid robot market reaches 18,000 units shipped in 2025, with a 1B-unit cumulative deployment by 2050 — implying the steepest growth ramp begins 2028-2030.[^17] Morgan Stanley modeled a $5 trillion aggregate market opportunity by 2050.[^15]
Practical takeaway: 2028 is when "humanoid robot leasing" likely becomes a mainstream offering — Apptronik, Figure, and Agility have all signaled robot-as-a-service business models in their public communications. Monthly cost per humanoid in the $1,500-$4,000 USD range is the analyst consensus. For an SMB in logistics, food service, or facilities management, the TCO crosses the human-FTE-equivalent line for night/weekend coverage in 2028, not 2027.
Year 3 (calendar 2029): vertical specialization
Cited driver: By 2029, you should expect purpose-tuned humanoids: a healthcare humanoid certified for hospital corridors (not patient contact), a retail humanoid optimized for shelf-stocking, an agriculture humanoid that operates on uneven outdoor terrain. The pattern follows software AI's path — the 2026 generalist Claude/GPT-5 spawned vertical-tuned models for legal (Harvey), medical (DAX), financial (Bridgewater's internal model). Robotics will follow the same fork.
Analyst projection: Citi Group's GPS report on Generative AI (June 2023) and follow-up coverage projected AI-augmented physical labor automation reaching 30% of warehouse roles by 2030.[^18] Goldman Sachs' 2024 humanoid update explicitly modeled vertical specialization beginning in the 2028-2030 window as the unit economics permit purpose-built variants.[^14]
Practical takeaway: A solo restaurateur, dental practice, or specialty retailer in 2029 should expect at least one credible humanoid-leasing vendor pitching a vertical-tuned solution at sub-$2,000/month operating cost. The economic decision shifts from "experimental pilot" to "standard back-of-house investment."
Year 4 (calendar 2030): mainstream deployment
Cited driver: The 4-year compound: model gains (4 generations of GR00T / π / Skild / DeepMind Gemini Robotics), unit-cost reduction (manufacturing scale + battery/actuator price curves following established consumer-electronics patterns), and a maturing regulatory/insurance/leasing ecosystem.
Analyst projection: Goldman Sachs: 1.4M units shipped annually by 2035, with the inflection visible by 2030.[^14] Bank of America: cumulative deployment scales aggressively from 2030 onward toward the 1B unit mark by 2050.[^17] Gartner's strategic-planning predictions for mid-decade enterprise AI place autonomous physical systems as a top-three investment category for 2030-bound CIOs.[^19]
Practical takeaway: 2030 is when the conversation shifts from "should I deploy a humanoid robot?" to "which one and how many?" for SMB operators in physical-work-heavy verticals. Software AI employees by then will have crossed the multi-hour-to-multi-day autonomous task horizon that METR's research trajectory projects (doubling roughly every seven months extrapolated[^20]). The SMB stack of 2030 looks like: a frontier model subscription + 2–4 software AI agents + 1–3 leased humanoid robots, all coordinated through the same agent-orchestration layer.
How to position your business now
Three honest, action-oriented postures for 2026 operators planning for the 2027–2030 arc:
-
If you run a digital-only business (SaaS, agency, content, consultancy): Software AI employees compose now. Robots are not your immediate planning horizon. Re-read Part 2's SMB stacks and start there. The trajectory will pull you forward; you don't have to chase it.
-
If you run a physical-work-heavy business (logistics, warehousing, restaurant, retail, healthcare administration, light manufacturing, facilities, agriculture): Plan a 3-year capital allocation review. 2027 is for partnerships and pilots, 2028 is for the leasing-model decision, 2029 is for vertical-specialized deployment, 2030 is for scale. Don't lease a humanoid in 2026 unless you're a public-relations beneficiary; the economics aren't there yet outside of automotive-grade plants.
-
If you're a solo operator or knowledge worker: The compounding software-AI-employee curve hits your productivity surface every quarter. The robotic curve hits in five years for the world's productivity. Plan for both, deploy the first today, watch the second.
The best general framework for thinking about which assets hold value through this transition — durable real assets vs. depreciating skills vs. equity in the labor-displacement winners — is in The Resale Trap. The book is about consumer goods (cars, condos, collectibles), but the framework — which costs go down vs. which incomes go down — is the same problem space the AI-and-robotics transition presents to a household balance sheet.
What could break the forecast
Three honest unknowns. None of these is "AI winter is back" — that's not what serious analysts or capital allocators are pricing in. The real risks:
-
Battery + actuator cost reductions are slower than the model-side curve. Hardware doesn't have a Moore's-Law-equivalent. If actuator cost stays flat while model capability soars, the cost-per-task-completed curve flattens regardless of how smart the brain gets. Watch unit-cost-per-humanoid in vendor disclosures.
-
Regulatory friction (US OSHA, EU Machinery Regulation, China State Council). Industrial-floor deployment is regulated. A serious incident in a high-profile pilot (someone hurt, public-safety event) compresses the deployment timeline by 6-18 months across the industry. Watch incident data and regulatory commentary.
-
Tail-risk in the foundation-model substrate. A capability plateau on the model side (the same models for 18 months without major capability gains) cascades into both software and robotics. The Klarna walk-back of 2025 is a small warning shot — quality on edge cases didn't keep up with marketing claims. The bigger version of that, at scale, would slow everyone.
The base case in the bank forecasts already prices these in. The optimistic case probably ships earlier; the pessimistic case probably ships 2031–2032 at the same scale rather than 2030.
Related reading
- Part 1: AI Employees in 2026 — capabilities by role and 2027 roadmap — the survey article this series began with.
- Part 2: Small-business AI stacks — the SMB-specific deployment guide.
- Part 4: Wider wave — legal / medical / hospitality / housekeeping / government + state impact + W-2 playbook — licensing moats, USG-protected work, average-person preparation checklist.
- Part 5: Personal robotics — cooking / driving / errands / pilots / trades — consumer-facing physical AI through 2030.
- The Agent Protocol Stack — MCP, A2A, and the wiring layer that connects agents to systems.
- AI model routing — when to use which model for which agent role.
Fact-check notes and sources
[^1]: NVIDIA, "GR00T: A foundation model for humanoid robots" announced at GTC March 2024. https://nvidianews.nvidia.com/news/foundation-model-isaac-robotics-platform
[^2]: Reuters / TechCrunch coverage, "Physical Intelligence raises $400M at $2.4B valuation" (November 2024). https://www.reuters.com/technology/artificial-intelligence/physical-intelligence-raises-400-million-bezos-openai-others-2024-11-04/
[^3]: TechCrunch, "Skild AI raises $300M to build a generalist brain for robots" (July 2024). https://techcrunch.com/2024/07/
[^4]: Figure AI press release / Bloomberg coverage, $675M Series B at $2.6B valuation, February 2024. https://www.figure.ai/news
[^5]: Figure AI / BMW partnership announcement (January 2024) and subsequent pilot demonstrations. https://www.figure.ai/news/bmw-manufacturing
[^6]: Agility Robotics commercial-deployment announcements (Amazon 2023, GXO Logistics 2024). https://www.agilityrobotics.com/news
[^7]: Apptronik Apollo announcements: Mercedes-Benz (March 2024), GXO multi-year contract (October 2024). https://apptronik.com/news
[^8]: Tesla AI Day demonstrations and Q4 2024 earnings call commentary on Optimus. https://www.tesla.com/AI
[^9]: Boston Dynamics, "An Electric New Era for Atlas" announcement (April 2024). https://bostondynamics.com/blog/electric-new-era-for-atlas/
[^10]: 1X Technologies Series B announcement, January 2024. https://www.1x.tech/discover
[^11]: Unitree Robotics product pricing pages and commercial release coverage. https://www.unitree.com/
[^12]: Physical Intelligence π0 model release blog post (April 2024). https://www.physicalintelligence.company/blog/pi0
[^13]: Google DeepMind, "Gemini Robotics" announcements covering vision-language-action models. https://deepmind.google/discover/blog/
[^14]: Goldman Sachs Research, "The Humanoid Robot Market Could Be Larger Than We Thought" (January 2024). https://www.goldmansachs.com/insights/articles/the-humanoid-robot-market-could-be-bigger-than-we-thought
[^15]: Morgan Stanley Research, "Humanoid Robots: A $5 Trillion Market by 2050" coverage and updates (April 2024). https://www.morganstanley.com/ideas/humanoid-robots
[^16]: EU Machinery Regulation 2023/1230 (in force from January 2027). https://eur-lex.europa.eu/eli/reg/2023/1230/oj
[^17]: Bank of America Global Research humanoid-robot forecasts, 2024 publications. Coverage: https://www.bofaml.com/
[^18]: Citi Group "Generative AI: Growth Without Limits" GPS report and follow-on commentary (2023-2024). https://www.citigroup.com/global/insights
[^19]: Gartner annual "Top Strategic Technology Trends" reports, 2024-2025 editions citing autonomous physical systems. https://www.gartner.com/en/articles/
[^20]: METR, "Measuring AI Ability to Complete Long Tasks" (March 2025). https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/
This post is informational — not investment, employment, or strategic advice. All forward-looking dates and dollar values are vendor projections, analyst estimates, or extrapolated from cited research; actual outcomes will differ. Mentions of third-party companies are nominative fair use; no affiliation, endorsement, or partnership is implied.
