Topics Pioneers Insight Method Author
Back

What iPhone Packaging Reveals About the Robotics Opportunity

·The a16z Show

Rendering the long image…

Topics Embodied AI and Humanoid Mass Commercialization
Pioneers Josh Gruenstein

Summary

  • The packaging around an iPhone exposes a largely manual industrial market. Gruenstein says nine out of ten US factories and warehouses have no robots, with 38,000 industrial robot work cells deployed in 2025. The elaborate automated factories familiar from television represent, in his telling, the “top 0.01%”; even iPhone packaging involves people assembling it by hand.
  • Full-stack control is Tutor’s answer to an immature robotics market. Gruenstein argues that excluding one layer removes “90% of your trade-off space,” making models, hardware, software, supply chain, customers and go-to-market inseparable. Its Cassie and Sunny robots “can probably do 80% to 90% of the work” within manufacturing and logistics, but his longer-term platform thesis allows renewed specialization as the market matures.
  • Paid robot work is Tutor’s proposed data moat, and full autonomy remains an economic choice. Customer payments establish that collected experience has commercial value; Gruenstein treats human tutors as another cost to optimize. Raghuram asks whether solving 90% could still leave 10% requiring humans: “That might be fine,” Gruenstein replies. Tutor’s 100-Sunny Data Factory One tests real-world reinforcement learning at fleet scale.
  • The sales pitch is “a robot for 16 bucks an hour,” without upfront customer CapEx. Gruenstein reports 40–60 days from first call to contract, about 40 more days to deployment, and production within a day, often hours. Generally capable, non-customized robots and an aligned business model make deployment a different proposition from commissioning a custom automation project.
  • Operating costs now exceed upfront hardware costs in Tutor’s economics. Pressed by Casado, Gruenstein says costs are “more ongoing at this point,” including inference, teleoperation, maintenance and break-fix alongside hardware amortization. Customers facing aging workforces and thin margins view adoption as “strategic,” making continuing cost and reliability central to the proposition.
  • Small cost improvements could change which production methods become viable. Gruenstein recounts factory-built houses costing about 5% more than equivalent conventional construction; if they became just 1% cheaper, he says developer demand would shift entirely toward factory production. His broader thesis is that an “infinite flexible labor substrate” could enable more factories and physical businesses.
  • Tutor expects deeper US manufacturing integration within an interdependent global supply chain. As its orders become a material share of suppliers’ output, OEMs become “glorified contract manufacturers,” changing the case for bringing components in-house. Tutor manufactures in Watertown, Massachusetts, and expects more domestic component production, while Gruenstein stresses that supply chains remain “highly interconnected” and “interdependent” in 2026.

Deep dive

1. Robotics is a systems business before it becomes a platform

  • Founded in 2021, Tutor Intelligence now has about 160 people, primarily in Boston, deploying robots while building models, datasets and manufacturing capacity. Gruenstein relays an estimate Claude gave him the previous night: Tutor robots have helped manufacture and ship “over 250 million products” to American consumers.

  • Raghuram asks whether the harder challenge is operations and unit economics; Casado adds go-to-market to the list. Gruenstein answers “No,” but says you have to do all of it. Even infinitely available intelligence leaves manufacturing, software and customers unsolved. Full stack means models, hardware and software plus supply chain, customers, product and business model. Refusing to touch one layer means “you’ve eliminated 90% of your trade-off space to make progress.”

  • Raghuram asks whether the physical world lacks the clean abstractions that helped horizontalize computing; Casado adds the challenge of navigating physical space. Gruenstein attributes today’s integration to market immaturity. A humanoid supplies a CPU-like general abstraction, he argues, while different embodiments can require different components, as computing uses CPUs, GPUs and ASICs. His expected sequence is specialized machines, general-purpose robots, then renewed specialization when the market becomes large enough.

2. Paid work turns robot deployment into a learning loop

  • Gruenstein’s coffee-cup example explains the missing dataset: picking up a cup leaves no internet record of fingertip pressures or poses, unlike email and social posts. Tutor connects deployed robots to human teleoperators, its “tutors,” who teach tasks and capture experience. The intended loop creates useful physical work while building an “internet of robot training data,” instead of trying to produce intelligence from effectively no data.

  • Raghuram challenges the generative-AI analogy: image and text training distributions match products people buy, while useful robot tasks might lie just outside the training distribution. Gruenstein contrasts head-mounted cameras sent around Indonesia with robots doing paid customer work: the latter supplies a market signal that the data has economic value. “If you can make money as a business,” he says, “then you’re probably also making data as a business.”

  • Raghuram’s second objection is the long tail: solving 90% might leave 10% requiring humans for a long time. “That might be fine,” Gruenstein replies. Tutors are a utilization and unit-economics question, like serving models on an H200: weigh operational complexity and cost against engineering “that last nine” of reliability. “We’re not purists”; simulation and internet-scale data can also widen coverage.

  • Asked about models like Astra, Gruenstein calls spatial reasoning a “sea change.” Labs’ commercially motivated post-training on CAD and 3D modeling connects frontier-model knowledge to action diffusion transformers; robotics has been “catapulted many years ahead in the future.” Tutor began with convolutional neural networks in 2021, so architectural changes are familiar. The durable asset, in his view, is “the flow of data” and the deployment loops it accelerates.

3. iPhone packaging exposes the gap between factories and automation

  • Gruenstein says How It’s Made creates a distorted picture: its elaborate automated factories represent, in his telling, the “top 0.01%” of American manufacturing and logistics. Million-square-foot facilities still contain people working at conveyors and folding tables. Even iPhone packaging is assembled by people. For backbreaking handling work, he cites three-month turnover and an average worker age of 55, with those workers nearing retirement.

  • Cassie addresses heavy handling with one arm, many suction cups and foundation models in the “200 million to 400 million parameter range.” Those models predict affordances and grasps that the hardware executes. Gruenstein ties the design to goods traveling on pallets or in cardboard boxes, creating extensive manual case and pallet handling. Hardware, model capability and customer work were designed together.

  • Sunny is an “industrial semi-humanoid” designed around three to ten years of factory operation at human rates and reliability. Wheels suit polished industrial floors; higher payload, faster movement and industrial collaborative safety standards support work alongside people. A growing external fleet performs more dexterous, two-handed manipulation. Gruenstein frames it as the design changes needed to turn a Unitree-like concept into durable factory equipment.

  • Within manufacturing and logistics, Gruenstein says Cassie and Sunny today “can probably do 80% to 90% of the work.” Their hardware cannot handle gravel, leaving construction and landfill-material recovery for future embodiments with legs. Home robots face technical and economic constraints: current designs are too industrial, and household willingness to pay does not match the value available in a factory. Nevertheless, “we’ll go anywhere.”

4. $16 an hour changes how customers buy and adopt robots

  • With nine out of ten US factories and warehouses lacking robots, Tutor opens with “a robot for 16 bucks an hour.” Gruenstein reports 40–60 days from first call to contract, about 40 more days to working on-site, and production within a day, often hours. Shipping generally capable robots without customer CapEx or custom engineering lowers the commitment required. He attributes distribution speed to that alignment of incentives and to putting no economic risk on the customer.

  • Raghuram asks why so many customers are local businesses rather than Fortune 50 experimenters. Gruenstein says small and mid-market companies constitute most of the market, and customer diversity forces a general product. His cautionary sequence: secure one large customer, spend two years engineering a bespoke system, then discover it fails or the customer’s problem has changed. He watched that pattern recur “five to 10 times.”

  • Casado asks whether robots become coworkers. Gruenstein distinguishes the factory-manager buyer from the worker using the robot daily. “A robot is a robot unless it works, in which case it’s a machine,” he says, quoting a tweet whose author he cannot recall. For workers, naming builds attachment; some customers name machines after their children, and repeat purchasers can name new robots.

  • Owning manufacturing lets Tutor push changes to its process every two weeks. Code ships daily, with major releases every two weeks; product, customer success and implementation operate as one team close to customers. Research can pull levers to get specific customer data and take moonshots, co-designing products around modeling breakthroughs. Gruenstein wants physical-product iteration to resemble software iteration, putting newly enabled capabilities into the field quickly.

5. Operating economics determine where robots create new markets

  • Gruenstein’s value equation is work delivered divided by cost. Casado presses on what cost includes and which component dominates; after initially saying Tutor drives all costs down, Gruenstein specifies that ongoing costs now exceed upfront costs. Inference, teleoperation, maintenance and break-fix count alongside hardware amortization. The old framing of a capital purchase recovered over ten years gives way to continuing operating economics.

  • Raghuram asks whether purchases cover Black Friday demand or belong to labor planning. Gruenstein calls them “strategic”: aging workforces, rising demand and higher fulfillment costs threaten businesses with little margin to absorb pressure. Customers want a long-term robotics partner while retaining expertise in their own products. His example is deodorant manufacturing: the customer handles deodorant; Tutor must handle robots.

  • Gruenstein’s discontinuity example comes from someone who worked in a housing factory, he thinks in Wisconsin. Factory-built homes were about 5% more expensive than equivalent conventional construction. If they became just 1% cheaper, he says, all developer demand would shift to factory production. His hope is a robot labor substrate that lets other businesses transform production; cheaper goods can expand consumption, and materials recovery could cross similar viability thresholds.

6. Scale pulls manufacturing and research into the platform

  • Gruenstein cites 38,000 US industrial robot work cells deployed in 2025. Casado asks whether that figure excludes FANUC and other old-world robots; Gruenstein’s brief reply does not clearly settle the scope. He adds that Unitree robots would have appeared in those numbers, but there were probably zero US deployments. The existing robot supply chain is small and, he argues, “wrong”: FANUC and ABB robots are mostly not made in the United States and are designed for systems integrators and fast, repetitive movements, rather than co-design with frontier models. Tutor builds robot bodies in Watertown, Massachusetts, while sourcing components such as actuators and cameras from OEMs with greater manufacturing scale.

  • Volume changes supplier relationships: once Tutor represents a material share of an OEM’s production, Gruenstein says the supplier becomes a “glorified contract manufacturer.” Specifying components versus manufacturing internally becomes a different decision. He expects continued vertical integration and more US-made componentry, arguing that Tutor needs a mature domestic robotics supply chain and has both an “obligation” and an “opportunity” to help build it.

  • On China, Gruenstein stresses that “today in 2026” supply chains remain “highly interconnected” and “interdependent,” with Chinese companies investing in the United States and US companies investing in China. He also calls US greenfield automation “a great equalizer”: China’s earlier automation investments might become “an enormous mistake” because it will need to tear systems down. If flexible robot labor becomes available, he expects more factories and easier formation of physical businesses, analogous to cloud software lowering barriers for digital businesses.

  • Data Factory One is Tutor’s 100-Sunny fleet in Boston, collecting internal data and testing real-world reinforcement learning. Gruenstein calls it, “to our knowledge,” the largest semi-humanoid data-collection farm in the United States. His rationale is scale: reinforcement learning “doesn’t really work at small batch sizes,” while a large fleet makes previously impractical experiments accessible. Real-world reinforcement-learning traces sit at the top of his data hierarchy.

Verification Notes

The transcript’s exchange does not clearly establish whether the 38,000 work-cell figure includes FANUC and other legacy robots.