Firmus, HappyRobot, and Eliyan just reminded investors AI needs picks and shovels

The model layer still gets the memes. The checks in early August keep finding the boring layers underneath. Firmus, the Australian AI factory and grid-aware infrastructure company, said it secured full commitments for a USD $2 billion strategic equity round, with post-money valuation now above $10.5 billion. Coatue and NVIDIA followed on. Blackstone and Jane Street joined. HappyRobot raised $150 million Series C around August 4 at a $1.2 billion valuation for agentic AI aimed at enterprise operations and logistics. Eliyan closed $145 million Series C around July 30 at a $1 billion valuation for AI chip interconnects, with a Seligman-led round and participation that coverage ties to Cisco Investments and Lumentum. This is picks and shovels money, not another foundation-model bakeoff. See the startups lane for the pattern.

What is an AI factory round actually buying?

Firmus Grid Limited, trading as Firmus Technologies, sells the unsexy constraint stack. Power, cooling, NVIDIA-centered AI factory buildouts, and cloud capacity across Australia and Asia-Pacific. A fully subscribed $2 billion strategic equity round at above $10.5 billion post-money is a bet that GPU clusters are still gated by electrons and real estate as much as by model weights. You can announce a frontier model from a stage. You cannot magic a substation into existence on the same timeline.

Coatue and NVIDIA following on is the industrial signal. Blackstone's tactical vehicles and Jane Street participating says the buyer set now includes infrastructure capital and trading firms that understand power markets, not only classic SaaS growth funds. Firmus says the raise pushes total new equity over the past year past $3 billion and funds Project Southgate-style expansion. That is balance-sheet language for more sites, more megawatts, and more dependency on execution in physical space.

If your funding thesis still starts with who has the best chatbot demo, update the spreadsheet. The scarce assets are megawatts, interconnect, and operational automation around the fleet. Regions that can pair renewable or surplus power with political willingness to host AI factories will keep pulling disproportionate capital, which is why Australia and Asia-Pacific keep showing up in Firmus's story.

There is also a geopolitical flavor investors prefer not to put in the headline. Diversifying AI factory geography away from a handful of U.S. and Northern Europe campuses is a risk trade as much as a growth trade. Firmus is selling capacity in places that want the jobs and can tolerate the grid load, which is a product category now.

HappyRobot and Eliyan fill different layers of the same thesis

HappyRobot's $150 million Series C at $1.2 billion valuation is the software automation layer. Tech.eu describes agentic AI for enterprise operations, with logistics and ops workflows as the wedge and Prysm plus Eurazeo among the names attached to the round in that coverage. The product story is agents that do coordination work humans hate, not another general chat surface. Dispatch, exception handling, vendor ping-pong. The ugly workflows that never made it into the first wave of copilots.

Eliyan's $145 million Series C at a $1 billion valuation is the silicon plumbing layer. Tech Funding News frames the company around AI chip interconnects, with Seligman leading and strategic participation linked to Cisco Investments and Lumentum. Training clusters die on bandwidth and packaging constraints long before they die on missing PowerPoint roadmaps. Optical and advanced interconnect stories keep getting funded because the alternative is wasting GPUs that cannot talk to each other fast enough.

Read the three rounds together. Energy-aware AI factories. Agents that run the business processes around those factories and supply chains. Interconnect silicon that keeps accelerators fed. That is a coherent stack bet. It is also a bet against the idea that the only company that matters is the one with the largest parameter count this month.

HappyRobot especially sits downstream of Firmus-class buildouts. Someone has to operate the messy human systems that move freight, schedule maintenance windows, and keep enterprise SLAs while the clusters hum. Eliyan sits upstream of both, deciding whether the silicon can actually saturate. Capital is drawing a vertical line through the stack instead of spraying only at model brands.

I also notice who is writing the checks. NVIDIA following into Firmus is industrial self-interest dressed as venture. Cisco-adjacent and optics-adjacent names around Eliyan make sense if interconnect is the bottleneck they sell into. Eurazeo and Prysm showing up on HappyRobot is a European ops-software bet with enterprise sales muscle implied. The LP mix tells you which layer of the stack each firm thinks is scarce.

What can still go wrong after the victory lap

AI factory platforms live and die on interconnection queues, transformer lead times, and local politics around water and noise. A fully subscribed round does not shorten a utility study. If Project Southgate-class sites slip, the valuation math starts arguing with the construction calendar.

Ops-agent companies die a quieter death. They get stuck in pilot purgatory where every logistics customer wants a custom workflow and refuses to pay platform prices. HappyRobot's Series C buys time to productize those edges. It does not abolish them.

Interconnect startups face design-win timelines that laugh at Twitter cycles. A billion-dollar label can arrive years before volume silicon clears qualification in a hyperscaler BOM. Eliyan's strategic investors help that path. They do not guarantee it.

Founders reading this should steal the stack map, not the valuation envy. If your company sits between power and models, or between models and paid ops workflows, you are closer to where August capital actually concentrated.

LPs should demand the same map in diligence. Ask how many megawatts are contracted versus aspirational. Ask what percentage of agent revenue is recurring production workflow versus services-heavy pilots. Ask which hyperscaler or OEM design wins are signed versus slideware. Those questions separate shovel businesses from valuation tourism.

Why picks and shovels is the adult trade in 2026

Foundation models still matter. They also face brutal price competition and enormous capex. Investors hunting durable gross margins keep sliding down the stack toward power, networking, orchestration, and vertical agents with paid enterprise workflows. That slide is rational even if individual valuations are aggressive.

None of these valuations are destiny. A $10.5 billion post-money AI factory platform still has to deliver capacity customers will renew. A $1.2 billion ops-agent unicorn still has to win against incumbents and internal scripts. A $1 billion interconnect story still has to ship through real design wins. Megarounds buy runway and headlines. Customers buy delivered megawatts, working agents, and silicon that clears a board bring-up.

Compare this cluster with the hyperscaler earnings tape elsewhere in our cloud coverage. When Azure, AWS, and Google Cloud are still printing acceleration, the shovel makers have a demand story that is not purely narrative. When those prints cool, these valuations will look different overnight. That is the bargain of infra investing in a boom.

I am filing this trio as August's clearest reminder that AI capital is paying for infrastructure and workflow leverage, not only model brands. More startups coverage from Younes Bekrar will keep following the shovel makers.

Related reading on Skarvonix: our startups category, the authors directory, and more from Younes Bekrar.

Primary sources and further reading: Firmus $2B strategic equity announcement, Tech.eu on HappyRobot Series C, Tech Funding News on Eliyan Series C.