Together AI's $800M Raise, Striding AI's 'Physical AI' Launch, and South Korea's $584B Alliance

Three July 2026 developments define the capital and hardware escalation underlying the AI race. Together AI's $800M raise (at $3.5B valuation) validates open-source AI cloud infrastructure as a distinct investable market; Striding AI's stealth exit commercialises fleet-scale Physical AI — coordinated multi-robot intelligence trained on generalised world models; and South Korea's $584B public-private semiconductor mega-alliance — the largest industrial investment in any country's history — locks in Samsung and SK Hynix's dominance in High-Bandwidth Memory for the next decade.

---

🤖 Together AI — The Open-Source AI Cloud Infrastructure Market

Why Open-Source AI Infrastructure Is a Distinct Business

The AI model market has bifurcated into two distinct segments:

Together AI sits in the middle segment — providing the cloud infrastructure optimised specifically for running open-weight models, serving enterprises that want: (a) data privacy (weights and data stay in Together's managed environment, not OpenAI/Google's), (b) model flexibility (swap Llama for Gemma for DeepSeek without re-engineering), (c) cost reduction (Together AI's inference pricing is 60–80% cheaper than equivalent OpenAI API calls for comparable open models).

Together AI's Business Model and Competitive Position

Together AI vs Competitors (July 2026):

Together AI's proprietary inference engine — the technical moat: Together AI's "Kairos" inference engine achieves lower per-token costs than competitors through:

1. Continuous batching — rather than waiting for a full batch to assemble, Kairos dynamically slots new requests into in-flight GPU computation, improving GPU utilisation from ~55% to ~87%
2. Speculative decoding — a smaller "draft" model predicts 4–8 tokens ahead; the full model verifies in parallel rather than sequentially, reducing effective inference latency by ~30%
3. Kernel-level optimisation — custom CUDA kernels for specific open-model architectures (Llama, Mistral, Gemma attention patterns) rather than generic transformer kernels

The $800M round — terms and valuation:

• Valuation: $3.5B post-money (up from $1.25B at Series B in 2024)
• Lead investors: Sequoia Capital (led), Tiger Global, Greenoaks Capital
• Use of funds: 60% GPU cluster expansion (targeting 100,000 H100-equivalent GPUs by Q1 2027), 25% inference engine R&D, 15% international expansion (EU sovereign cloud compliance)

---

🦾 Striding AI — Fleet-Scale Physical AI Architecture

The Transfer Learning Problem in Robotics

The critical bottleneck in physical AI (intelligent robotics) has been the sim-to-real gap: AI models trained in simulation fail when deployed in real physical environments because:

• Simulated physics is perfect; real-world physics is noisy (friction variation, sensor noise, actuator backlash)
• Simulated visual environments are clean; real environments have varied lighting, occlusion, and reflections
• Single-task training overfits: a robot trained to pick a red block fails with a blue block, then fails with a cube

Striding AI's approach — generalised world model for robotics:

Target markets and initial deployments:

Funding at launch: $95M Series A led by Lux Capital + a16z; $40M strategic investment from Hyundai Motor Group (wanting fleet robotics for their automobile manufacturing plants).

---

🇰🇷 South Korea's $584B Semiconductor Alliance — HBM Dominance

Why High-Bandwidth Memory Is the Critical AI Hardware Chokepoint

LLM inference requires moving model weights from memory to compute at high speed. The key bottleneck is not floating-point operations per second (FLOPS) — it is memory bandwidth. High-Bandwidth Memory (HBM) is the only memory technology with sufficient bandwidth to feed modern AI accelerators:

Memory technology bandwidth comparison:

South Korea's HBM market dominance:

• SK Hynix: 53% of global HBM production (dominant supplier for Nvidia H100/H200/B200)
• Samsung: 38% of global HBM (growing; recently qualified for Nvidia's supply chain after yield improvements)
• Micron (US): 9% (small but growing, focus on HBM3e for AMD MI300X)

The $584B Alliance — Structure and Goals

Capital allocation (President Lee Jae Myung's announcement):

The "triple axis" strategic rationale: President Lee framed the alliance around three interdependent national priorities:

1. Semiconductors — HBM supply chain; foundry capacity in non-TSMC risk mitigation
2. Physical AI — domestic robot manufacturing powered by Korean HBM and AI chips
3. Green energy data centres — Korean-built, Korean-powered AI infrastructure for sovereign cloud

Geopolitical context: South Korea's alliance is a direct response to the US CHIPS Act ($52.7B), Taiwan's chip sovereignty investments, and China's own $41B National Integrated Circuit Industry Investment Fund III. In the event of cross-strait military conflict affecting TSMC, South Korean fabs are the only Western-aligned alternative with sufficient scale.

---

📌 The Bottom Line

• together-ai-800m-open-source-cloud: $800M at $3.5B valuation (from $1.25B in 2024); Kairos inference engine: continuous batching (55%→87% GPU utilisation), speculative decoding (-30% latency), custom CUDA kernels; Llama 4 405B: $3.50/1M tokens (vs OpenAI $15 equivalent); 60% GPU capex ($100K H100-equivalent by Q1 2027), 25% R&D, 15% EU expansion; enterprise migration driver: 60-80% cheaper + data privacy + model swap flexibility vs closed-source APIs.
• striding-ai-fleet-robotics: $95M Series A (Lux + a16z) + $40M Hyundai strategic; multi-modal world model: RGB/depth + proprioception + force/torque fusion transformer; 500M+ real-robot-hours training data (no simulation dependency); decentralised fleet coordination (10-1,000 robots, no central server); continuous learning: individual robots share policies without raw data (privacy-preserving); targets: logistics (100-5,000 AMRs), construction (legged robots), mining, manufacturing.
• south-korea-584b-semiconductor: $584B total: $220B 4 new SW-region fabs, $85B CoWoS-equivalent HBM packaging, $120B HBM4/4e R&D, $95B green AI data centres, $34B talent (100K engineers/10 years), $30B infrastructure; SK Hynix 53% + Samsung 38% = 91% global HBM share; HBM bandwidth: DDR5 89GB/s → HBM3e 1,200GB/s → HBM4 2,000GB/s; "triple axis": semiconductors + physical AI + green data centres; TSMC contingency: only Western-aligned alternative at scale in Taiwan-conflict scenario.

---

📬 Stay Updated

Get the best of AI & technology delivered to your inbox every week. Subscribe to our free newsletter →

---

Disclosure: This post contains affiliate links. If you purchase through our links, we earn a small commission at no extra cost to you. We only recommend products we believe in.