Supply Chain Optimization

Draw the physical chain end to end, and refuse to stop at tier one

For a single automotive microcontroller the chain is: silicon wafer maker, foundry (a Renesas fab such as Naka, or TSMC), then assembly-test-and-packaging houses concentrated in Malaysia and the Philippines, then the chip company that sells the finished part, then a tier-one module maker like Denso, Aisin or Continental, then Toyota's assembly plant, then the dealer. Toyota's purchase order goes to Denso for a body-control module. The node that failed sat three tiers upstream and did not appear on any Toyota contract, which is exactly the lesson Toyota had already learned after the 2011 Tohoku earthquake, when it could not name who made the chip inside a wiper motor. Drawing the chain past your direct supplier is the whole game here; a candidate who stops at Denso never finds the constraint.

Attach a cost and a time to every node

An automotive-grade microcontroller takes roughly 26 weeks of irreducible physical process to produce: about 12 to 16 weeks in the fab, plus test, qualification and packaging. Industry lead times stretched from around 13 weeks before the pandemic to over 20 weeks by late 2021, with some automotive MCUs quoted close to a year (approximate industry figures). Now the killer ratio: semiconductor content is roughly 500 dollars on a vehicle with a 30,000 dollar average price, so about 2 percent of the bill of materials gates 100 percent of output. Toyota's own assembly node was never the problem, it holds about two hours of parts at the line and ran fine; the constraint was a cheap, small, long-lead, single-sourced component.

Separate the demand-signal problem from the physical-flow problem

Automakers cut chip orders hard in Q2 2020 on a forecast that vehicle demand had collapsed, then demand rebounded within months and they tried to reinstate orders. Foundries had already reallocated that wafer capacity to consumer electronics, which was booming, orders in larger volumes and pays more per wafer. Autos are only around 8 to 10 percent of global semiconductor demand, so the industry is not the priority customer when capacity is scarce. This is a textbook bullwhip amplified by a capacity-allocation reality, and it means no amount of plant scheduling, logistics or warehouse work would have fixed it.

Find the root cause of Toyota's difference, which is a deliberate purchase of resilience

After 2011 Toyota built a multi-tier supplier database mapping parts several levels upstream, and required suppliers to hold buffer stock of roughly two to six months on about 1,200 risk-flagged parts, semiconductors among them (as reported at the time, approximate). That is not a violation of just-in-time; it is just-in-time applied selectively, with the buffer bought precisely where the part is cheap, non-perishable, long-lead and capable of stopping the entire line. Toyota also had the multi-tier map, so when the Renesas Naka fab caught fire in March 2021 it could see which of its own vehicles were exposed within days rather than weeks. The competitive advantage was information first and inventory second.

Size the prize, on contribution and not revenue

An unbuilt vehicle is not lost revenue, it is lost contribution, so use roughly 15 to 20 percent of a 30,000 dollar average price, about 5,000 dollars a unit (illustrative). Toyota's September 2021 cut of roughly 360,000 units below plan therefore put on the order of 1.8 billion dollars of contribution at risk in a single month. Now price the buffer that avoided that for the earlier months: three months of chips at 500 dollars a car on a 9 million unit annual build is about 1.1 billion dollars of inventory, costing perhaps 90 million dollars a year at an 8 percent cost of capital plus some obsolescence risk. Paying roughly 90 to 150 million a year to protect something worth 1.8 billion a month is not a close call, and stating it in those terms is what turns an operations observation into a board-level recommendation.

Generate levers and name what each one costs

Four levers, each with its price. One, multi-tier mapping and risk-scoring parts on criticality times lead time times single-sourcing, then buffering only the top decile: cheap, but it requires suppliers to disclose their own suppliers, which they resist. Two, chip commonisation across platforms so one MCU serves many models: cuts variety and raises the pooled buffer's efficiency, but slows feature differentiation and locks the design. Three, contract directly with the foundry instead of through the tier-one, which Toyota and Denso effectively did by backing TSMC's Japanese fab venture from late 2021: it buys allocation priority but is multi-billion-dollar capex with a long payback. Four, spec flexibility, building and shipping vehicles with a feature deferred and retrofitted later, which is fast and free but degrades the customer experience and creates a rework tail.

Recommend with sequencing and a risk

Immediate, no capex: extend the multi-tier map to every part with lead time above 12 weeks, and set buffer targets by risk score rather than uniformly, since buffering everything would destroy the working-capital advantage that justifies just-in-time in the first place. Medium term: commonise microcontrollers across platforms and dual-source at the fab level, not just at the tier-one level, because two tier-ones buying from the same fab is not dual sourcing. Long term and capex-heavy: equity or long-term-agreement positions in foundry capacity, justified only if the option value of allocation priority beats the cost of capital across a cycle. Biggest risk: buffers on chips age badly against 6 to 8 year model lifecycles and design refreshes, so the buffer policy needs an obsolescence review, and the one datapoint to demand is true sub-tier capacity by part, not the tier-one's assurance that supply is fine.

Draw the chain in both directions, because the reverse chain is where the money is

Forward: Surat fabric mills, job-work stitching units in Jaipur and Ahmedabad, inward QC at Bhiwandi, 3PL first mile, origin sort hub, line haul, destination hub, last-mile rider, customer. Reverse: a failed delivery or a customer return goes back to the destination hub, onto a return-to-origin lane, into Bhiwandi inward, through QC and re-grading, and then either back into sellable stock or into write-off. That reverse path has five nodes, takes 13 to 18 days, and typically has nobody on the org chart carrying a KPI for it, which is the SCOR Return process being run by accident. Say out loud that you will quantify both directions, because a candidate who only draws the forward arrow will diagnose this case as a marketing problem.

Attach cost and days to each node, then demand the RTO reason-code split

Forward freight is about Rs 65 a shipment and 4.5 days to a tier-2 or tier-3 pincode against 2.2 days to a metro; reverse freight is about Rs 60 and 13 to 18 days. RTO (return to origin, the parcel that never reached the customer) runs at 27 percent on COD orders and 4 percent on prepaid, and a further 12 percent of delivered orders come back as post-delivery returns. The single highest-value data request in this case is the RTO reason-code split, which comes back roughly as: customer unreachable or refused at the door 44 percent, incomplete or misrouted address 23 percent, courier attempt failed or falsely marked 19 percent, product or expectation mismatch 14 percent (illustrative). Three of those four are information failures at the point of order capture, not physical logistics failures, and that reframes the entire case.

Benchmark node by node and name the broken one

Make is fine: job-work first-pass yield is 96 percent and OTIF into Bhiwandi is 94 percent, both roughly at industry norm. Forward Deliver is acceptable: the 3PL hits its promised SLA on 92 percent of shipments. Prepaid RTO at 4 percent is normal for Indian fashion; COD RTO at 27 percent is at the ugly end of the typical 20 to 25 percent band for the category. So the broken nodes are the order-capture node, where an unverified phone number and a free-to-refuse payment method create a fake order, and the last-mile node, where courier performance varies enormously by pincode. Commit to that: the constraint is order quality, not warehouse throughput.

Size the prize with the fully loaded cost of an RTO, not the return freight

Of 20 lakh shipments, COD is 12.4 lakh with 27 percent RTO (3.35 lakh) and prepaid is 7.6 lakh with 4 percent RTO (0.3 lakh), so about 3.65 lakh RTOs, 18 percent of everything shipped. Per RTO: forward freight Rs 65, reverse freight Rs 60, packaging Rs 25, warehouse re-grading and QC Rs 15, so Rs 165 of hard logistics cost; add roughly 7 percent of returned units coming back soiled or damaged and written off at a Rs 640 cost, another Rs 45 on average; then add the Rs 240 of performance-marketing spend that acquired an order which never converted to a sale. That is about Rs 450 fully loaded per RTO, so 3.65 lakh times Rs 450 is roughly Rs 16 crore a year. Against net delivered revenue of about Rs 209 crore that is 7.6 percent of the top line, which on its own is larger than the EBITDA hole.

Check whether the demand signal and the network are also feeding the problem

Two secondary findings. First, buying is done as one seasonal drop with an 11-week reorder lead time from job-work units, so hits cannot be repeated and misses cannot be cancelled; 2,400 SKUs are live but 300 SKUs drive 60 percent of volume while the tail sits at 140 days of stock and is eventually dumped through discounting. Second, the single Bhiwandi node means a 4.5-day promise into tier-2 and tier-3 where competitors promise two, and a longer promise mechanically raises refusal and cancellation, which loops straight back into the RTO number. So the network is not the primary cause but it is an amplifier, and the SKU tail is a separate cash problem worth its own line.

Generate levers and kill the obvious one on the trade-off

The instinctive lever is a prepaid discount. Test it: moving COD share from 62 to 45 percent shifts about 3.4 lakh orders to prepaid, each avoiding 0.23 of an RTO, so about 78,000 fewer RTOs times Rs 450 is roughly Rs 3.5 crore saved, but a Rs 75 discount on 3.4 lakh orders costs Rs 2.55 crore, so the lever nearly breaks even. The boring levers win instead: a Rs 49 COD convenience fee is revenue-positive and shifts the same behaviour, mandatory phone-OTP and address validation at checkout kills much of the 23 percent address bucket, a WhatsApp confirm-before-dispatch on COD orders above Rs 2,000 costs well under Rs 1 crore across the base and attacks the 44 percent unreachable bucket, and courier allocation by pincode-level RTO scorecard rather than by rate card fixes the 19 percent failed-attempt bucket. Also make the worst 400 pincodes prepaid-only. Forward-deploying inventory into Delhi NCR and Bengaluru hubs for the top 300 SKUs would cut the promise to 2.5 days, but splitting stock across three locations raises safety stock by roughly 1.7 times by the square-root law and adds about Rs 1.2 crore of fixed cost, so it is justified only on the fast-moving 300, never on the tail.

Recommend with sequencing, a number and a risk

First 90 days, zero capex: COD fee, OTP and address validation, WhatsApp dispatch confirmation, pincode-level courier scorecards and prepaid-only for the worst pincodes. Target RTO from 27 to 18 percent on COD, which is roughly 1.1 lakh fewer RTOs and about Rs 5 crore, plus the COD fee revenue and lower write-offs, so call it Rs 8 to 9 crore of annualised benefit against under Rs 1 crore of cost. Next 6 to 12 months, structural: two forward hubs for the top 300 SKUs, and a second job-work partner on a four-week replenishment cycle so the fast movers can be chased and the tail can be cut from 2,400 to about 1,400 SKUs. Biggest risk: the COD fee suppresses conversion in exactly the tier-2 and tier-3 geographies driving growth, so A/B test it by pincode cluster and judge it on contribution per session rather than on conversion rate, and the one datapoint to pull before committing is delivered-order economics by pincode cluster rather than by channel.