Breaking the Bottleneck | Issue 63
[11/4/2024] Concrete, Manufacturing Policy, and New Foundation Models!
Breaking the Bottleneck is a weekly manufacturing technology newsletter with perspectives, interviews, news, funding announcements, manufacturing market maps, and a startup database!
💥 If you are building, operating, or investing in manufacturing, hardware, or robots, please reach out. My email is aditya@schematicventures.com – I’d love to connect!
🏭 If you were forwarded this and found it interesting, please sign up!
Content I Enjoyed Last Week 🏭🗞️🔬 📚
News:
Big Data Means Big Concrete [IEEE Spectrum]
The rapid expansion of data centers drives the demand for concrete, particularly cement, a major contributor to global greenhouse gas emissions - accounting for about 6% of global CO₂ emissions. Concrete's carbon footprint stems mainly from two processes in cement production: combustion heating limestone to around 1,500°C in kilns requires burning fossil fuels, generating 35-50% of the industry's emissions, and calcination, the transformation of limestone (calcium carbonate) into clinker releases CO₂ directly, accounting for most of the remaining emissions. To address these challenges, the concrete and cement industry is exploring several innovations:
Carbon Capture and Storage (CCS): Major cement producers like Holcim, Heidelberg Materials, and Cemex are testing CCS technology to capture CO₂ emissions and store it underground or use it in other products.
Alternative Cement Formulations, such as:
LC3 Cement: Developed by researchers at the Swiss Federal Institute of Technology Lausanne (EPFL), LC3 blends calcined clay and ground limestone, reducing emissions by 30-40% compared to Portland cement.
Fly Ash-Based Cements: Utilizing fly ash, a byproduct of coal-fired power plants, to replace a portion of cement in concrete mixes.
Recycling and Reusing Concrete: Companies are developing methods to recycle concrete from demolished structures, reducing the need for new production.
AI-Designed Concrete Mixes: Researchers are using artificial intelligence to create concrete mixes that require less cement while maintaining strength and durability. A collaboration between the University of Illinois, Ozinga, and Meta resulted in a concrete mix with a 40% lower carbon footprint for data centers.
How US Tax Breaks Brought a Chinese Solar Energy Giant to Ohio [Bloomberg]
Illuminate USA's new solar factory in Pataskala, Ohio, is bustling with activity as hundreds of newly hired local employees transform sheets of glass into state-of-the-art photovoltaic panels. On the surface, this appears to be a hallmark of the 21st-century clean energy manufacturing boom promised by the Biden administration. However, this success story also benefits China. Illuminate USA is a joint venture owned 51% by Invenergy and 49% by Longi Green Energy Technology Co., a Chinese solar giant. Longi provides the panel-making expertise, technology, and supply chain, allowing the production of tariff-free equipment for the U.S. market. By partnering this way, Longi avoids U.S. tariffs on Chinese solar panels and benefits from up to $350 million in potential annual tax subsidies from the Inflation Reduction Act (IRA). This strategy is not unique to Illuminate; companies based in or linked to China are planning to build at least a dozen plants in the U.S. with 30 gigawatts of module-making capacity, enough to supply roughly three-quarters of today's U.S. panel needs. This situation has created a unique case for the US government, where Invenergy can secure domestically produced panels for its solar arrays, and local workers see the job as a chance to double their income. Still, it creates a reliance on foreign investment/support, particularly China, for a critical industry moving forward.
Five Smart Policies Can Turbocharge Clean US Manufacturing [Canary Media]
The IRA and the Bipartisan Infrastructure Law are already yielding positive results. However, industrial firms emit one of every four tons of climate pollution. With increasing global demand for industrial products, the U.S. must implement five smart policies to drive investments in industrial innovation:
Enact a Tax Credit for Clean Industrial Heat: Since 85% of fossil fuels burned by industry produce heat for processes like melting metals and molding plastics, a clean heat production tax credit would incentivize manufacturers to switch to non-polluting energy sources.
Reform Electricity Markets to Value Flexible Energy Storage: The Federal Energy Regulatory Commission (FERC) and state public utility commissions should adjust electricity markets to value highly flexible energy-storage technologies. As industries transition to clean energy, they will demand more electricity, which can be supplied by utilizing capacity in existing power plants and batteries.
Reauthorize and Expand the DOE’s Industrial Demonstrations Program: This transformative program aimed at building first-of-a-kind clean manufacturing facilities and retrofitting existing plants will create tens of thousands of jobs and reduce climate pollution by 14 million metric tons annually.
Extend and Expand the Qualifying Advanced Energy Project Credit: This cross-cutting incentive supports multiple U.S. goals, including clean manufacturing and critical minerals development.
Implement a Carbon Border Tariff Based on Greenhouse Gas Emissions: Congress should enact a tariff on imported products based on the greenhouse gases emitted during manufacturing to protect domestic manufacturers, raise revenue for reinvestment, and incentivize foreign firms to access the U.S. market.
Supply Chains, Still Vulnerable [McKinsey]
According to the latest McKinsey Global Supply Chain Leader Survey, nine in ten respondents reported encountering supply chain challenges in 2024. Despite these issues, there are signs that companies are reducing their efforts to enhance supply chain resilience, potentially leaving them vulnerable to future disruptions. The survey reveals significant gaps in organizations' abilities to identify and mitigate supply chain risks. Some important highlights:
73% of survey respondents report progress on dual sourcing, and 60% are regionalizing their supply chains. The share of companies with comprehensive visibility of their tier-one suppliers has reached 60%, marking a consecutive annual increase of ten percentage points.
Two-thirds of surveyed companies are investing in advanced planning and -scheduling (APS) systems, and only 10% have completed their deployments. One-third of respondents lack quantified business cases for APS systems, and 15% report that implementations have not met business objectives.
The share of companies with good visibility into deeper supply chain levels has declined by seven percentage points for the second year.
Companies take an average of two weeks to plan and execute a response after a supply chain disruption.
Only 9% of companies report their supply chains are currently compliant, and 30% admit they are behind in compliance efforts.
Can AI Learn Physics from Sensor Data? [Archetype AI]
The Archetype AI Team has developed a physical AI foundation model named Newton, capable of encoding and predicting physical behaviors and processes it has never encountered before without being explicitly taught the underlying physical principles an via learning directly from data. They trained Newton on 0.59 billion samples from open-source datasets covering various physical behaviors, including electrical currents, river fluid flows, and optical sensors. Utilizing a transformer-based deep neural network, Newton encodes raw, noisy sensor data, uncovering hidden patterns and statistical distributions in a universal latent space representing these measurements. In experiments, Newton was tested on simple physics scenarios familiar from school experiments, such as mechanical oscillation and thermodynamics. Remarkably, without being specifically trained in these experiments, Newton accurately predicted the behavior of these physical systems—even for chaotic and complex behaviors—demonstrating zero-shot forecasting capabilities. The team then challenged Newton with complex real-world systems, including predicting city electrical demand, daily temperature variations, and oil temperature in electrical transformers. Newton accurately forecasted the behavior of these complex systems with no additional training data specific to those systems. Newton's zero-shot forecasting consistently outperformed models trained solely on the target datasets, suggesting that foundation models like Newton have powerful generalization capabilities, enabling them to understand physical behaviors far beyond the specific data they were initially trained on.
Our First Generalist Policy [Physical Intelligence]
Researchers at Physical Intelligence have developed a general-purpose robot foundation model called π0 (pi-zero), aiming to create artificial physical intelligence that allows users to simply instruct robots to perform any task—much like interacting with large language models (LLMs) and chatbot assistants. Over the past eight months, the team has trained π0 on broad and diverse data, enabling it to follow various text instructions and control different robots. Unlike LLMs, π0 spans images, text, and actions, acquiring physical intelligence by training on embodied experiences from robots and learning to output low-level motor commands through a novel architecture directly. π0 is trained on the largest robot interaction dataset to date, including both open-source data and a diverse dataset collected across eight distinct robots. The tasks in this dataset involve various motion primitives, objects, and scenes, covering activities such as Bussing dishes, Packing items into envelopes, Folding clothing, Routing cables, Assembling boxes, Plugging in power plugs, Packing food into to-go boxes, Picking up and disposing of trash. The model inherits semantic knowledge and visual understanding from Internet-scale pretraining, starting from a pre-trained vision-language model (VLM). The VLM is a smaller 3-billion-parameter model trained on text and images from the web. To enable high-frequency dexterous robot control, the team developed a novel method to augment pre-trained VLMs with continuous action outputs via flow matching, a variant of diffusion models. This allows π0 to output motor commands up to 50 times per second, essential for fine-grained manipulation tasks. While π0 can perform tasks zero-shot, more complex and dexterous tasks require fine-tuning with high-quality data. The team compared π0 to other robot foundation models and attained good performance across all tasks. In contrast, the smaller variant, π0-small (a 470-million-parameter model without VLM pre-training), achieved the second-best performance but with more than a 2x improvement when using the full-size architecture with VLM pre-training.
Research:
World Energy Outlook 2024 [IEA]
The IEA's flagship World Energy Outlook, published annually, is the most authoritative source of global energy analysis and projections. Some interesting takeaways:
Since 2020, almost 200 trade measures affecting clean energy technologies – most of them restrictive – have been introduced worldwide, compared with 40 in the preceding five-year period.
In 2023, renewables provided 30% of the global electricity supply, while fossil fuels edged down to 60%, their lowest share in 50 years. By 2035, the share of solar PV and wind in electricity generation will exceed 40% globally in the STEPS, and by 2050, it will increase to nearly 60%. The share of nuclear power remains close to 10% in all scenarios.
Over the past five years, annual solar capacity additions quadrupled to 425 GW. Still, yearly manufacturing capacity is set for a sixfold increase to more than 1 100 GW, a level that – if deployed in total – would be very close to the amounts needed in the NZE Scenario.
There are more than 11,000 data centers registered worldwide, and they are often spatially concentrated, so local effects on electricity markets can be substantial. However, at a global level, data centers account for a relatively small share of overall electricity demand growth to 2030
China stands out: it accounted for 60% of the new renewable capacity added worldwide in 2023 – and China’s solar PV generation alone is on course to exceed, by the early 2030s, the total electricity demand of the United States today.
Combining Next-Token Prediction and Diffusion in Robotics [MIT News]
Researchers from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) have developed a new sequence model training technique called "Diffusion Forcing" that enhances the flexibility and reliability of sequence modeling in artificial intelligence. Full-sequence diffusion models (e.g., Sora) generate entire sequences by successive denoising but cannot handle variable-length sequences. Diffusion Forcing bridges the gap between these models by introducing different noise levels to each token, effectively serving as a type of fractional masking. This allows the model to "unmask" tokens and generate sequences that are both future-aware and variable in length. Lead author Boyuan Chen, an EECS PhD student and CSAIL member, explains: "Sequence models aim to condition on the known past and predict the unknown future, a type of binary masking. However, masking doesn't need to be binary. With Diffusion Forcing, we add different levels of noise to each token... It knows what to trust within its data to overcome out-of-distribution inputs." Implemented in a robotic arm, it helped swap two toy fruits across three mats, requiring long-term planning and memory. The robot completed the tasks despite starting from random positions and facing visual distractions. The research team plans to scale up their method using larger datasets and advanced transformer models to improve performance. They aim to develop a ChatGPT-like "robot brain" that assists robots in performing tasks in new environments without human demonstration.
Podcasts/Video:
Democratizing Industrial Automation [Manufacturing Executive]
Manufacturing Deals🏭💵
Physical Intelligence - A company building foundation models for robotics and the physical world
$400 million [Venture] - Co-led by Thrive Capital and Lux Capital and joined by Bond, Jeff Bezos, OpenAI, and Redpoint Ventures.
Outrider - A company building self-driving tech for electric yard trucks
$62 million [Series D] - Co-led by Koch Disruptive Tech and NEA and joined by joined by 8VC, Ark Invest, B37 Ventures, FM Capital, Prologis Ventures, and others.
Andium - A company transforming wellsite management for oil and gas
$21.7 million [Series B] - Led by Aramco Ventures and joined by joined by Climate Investment, Intrepid Financial Partners
Third Wave Automation - A company building a shared autonomy platform for autonmous forklifts
$27 million [Series C] - Led by Woven Capital and joined by Innovation Endeavors, Norwest Venture Partners, and Qualcomm Ventures
Downtime 🏭🧑🔧
Dune Prophecy